Tech Lead · GenAI @ Elastiq AI 12.97°N / 77.59°E · Bangalore

GCP · Professional MLE --:--:-- IST

① Folio · 2026 ② Bangalore, IN ③ Technical Lead — GenAI @ Elastiq AI

Priyanshu Shekhar Sinha

Technical Lead — GenAI · AI/ML Engineer · GCP Professional MLE

gen.intro // streaming

Seven plus years building production-grade AI systems. Currently architecting an enterprise RAG platform at Elastiq AI — multi-source ingestion, hybrid search, cross-encoder re-ranking, and document-level access control.

Download Résumé ↗ View work → Get in touch →

Scroll

01 / About

Engineer of tangible outcomes.

Translating ambiguous business problems into deployed, measurable, data-driven systems.

I'm an AI/ML Engineer with seven plus years shipping production-grade data systems. I've grown from classical data science into the architecture of large-scale enterprise AI — currently leading the design and development of an enterprise RAG platform at Elastiq AI.

My work sits at the intersection of search infrastructure, LLM orchestration, and enterprise security — multi-source connectors (Drive, GCS, S3, Azure Blob, Slack), hybrid retrieval (BM25 + kNN on OpenSearch), cross-encoder re-ranking, and document-level access control synchronized across cloud providers.

I believe most of the work is finding the right shape of the problem. The rest is craft, iteration, and the discipline to keep what works.

Currently building: Enterprise RAG · ACL-aware retrieval
Lately exploring: Agentic AI · MCP connectors
Always for hire on: GenAI · LLM fine-tuning

0⁺

Years in practice

Projects shipped

0⁺

Verticals

0⁺

Mentored

02 / Stack

Tools, in daily use.

A pragmatic kit refined over a decade of building, breaking, and shipping.

Top Skills

GenAI Solution ArchitectAgentic AIEnterprise RAGGoogle Cloud Platform

Languages

PythonSQLCypher

LLM & GenAI

LLaMA 3.1 70BQLoRALoRASFTRLHFClassic RAGGraphRAGCypher-RAGText2SQLMCPLangChainTransformersOllama

Search & Retrieval

OpenSearchBM25kNN / VectorHybrid SearchCross-encoder Re-rankingChromaDBEmbeddingsACL Sync

Multi-modal

TextPDFAudioVideoSpeaker DiarizationOCR

ML / DL

PyTorchTensorFlowscikit-learnMLFlowPandasPolarsNumPyPySparkNLPDeep LearningMLOps

Cloud

GCPAWS S3AWS EKSAWS ECSAWS SageMakerAWS SQSAWS SESEC2GCSAzure Blob

Data

PostgreSQLCitus DBSnowflakeDatabricksGraph DBChromaDB

Tooling

DockerAirflowStreamlitChainLitFlaskPlotlyAltairPower BIPostmanGitGitHubGitLab

Soft Skills

CommunicationProject ManagementTeam LeadershipMentorship

02·B / Off-screen

When I'm away from the keyboard.

Long Drives

Badminton

Hiking

Cuisine

Documentaries

Music

Photography

Mentorship

03 / Echo

What collaborators say.

Selected notes from former managers, peers, and mentees.

Priyanshu got a very deep understanding of our pipeline and quickly recognized nuances in the Python-pandas framework. He learned and implemented ML models for our text-processing pipeline which improved accuracy and reduced false positives, and made significant contributions to our Learning Framework.

Venkatesh Mohan

Founder & Director — Sumyag

Priyanshu has immense knowledge — and not just that, he focuses on the core concepts. Conceptually it's very difficult to beat him. Whether or not he has worked on a certain algorithm, he picks it up in upcoming projects. His grasping power is amazing. He is a great asset to the team.

Sakshi Sehgal

Senior Data Scientist — Paxcomm

Priyanshu is highly focused and relentless. His incredibly deep knowledge of neural networks, deep learning, and ML could be an asset to any organization. He gels with the team well, always provides constructive feedback, and I would love to work with him again.

Garima Thakur

ML Engineer — Softuvo Solutions

Priyanshu is one of the most dependable, thorough, and dedicated people I've worked with. His contributions in document enrichment and image-processing initiatives improved customer experiences significantly. Priyanshu is an amazing value-add to any organisation.

Abhishek Kumar

Engineering Manager — Google

04 / Recognition

Selected honors.

Certifications, awards, and contributions worth flagging.

Google Cloud Cert · Active

Professional Machine Learning Engineer

Designing, building, and productionising ML solutions on Google Cloud — feature engineering, model serving, MLOps, and responsible AI.

Elastiq AI March · 2026

Star Employee of the Month

Excellence grows through continuous learning and self-mastery. — presented by the Elastiq leadership team.

Recruitment Smart 2021

Star Performer of the Month

Recognised for impact across reporting automation — reducing analytics report time from 7 days to 15 minutes (99% improvement).

GitHub Open Source

Arctic Code Vault Contributor

Open-source code preserved in the GitHub Archive Program — a 1,000-year repository archived in the Arctic.

GitHub Pull Shark

Pull Shark

Awarded for sustained contributions across multiple open-source repositories.

DeepLearning.AI Specializations · 2018, 2021

Deep Learning & NLP Specializations

Two multi-course Coursera specializations covering DNNs, CNNs, sequence models, attention, and the foundations of modern NLP.

05 / Résumé

Trajectory, in brief.

A linear walk through the past seven years, plus the studies that led there.

Education & Certifications

★ Cert

Google Cloud — Professional ML Engineer

Google Cloud

Designing, building, and productionising ML solutions on Google Cloud — feature engineering, model development, serving, MLOps, and responsible AI.

2021

NLP Specialization

DeepLearning.AI · Coursera

Classification & Vector Spaces · Probabilistic Models · CNNs · Sequence Models · Attention Models.

2019

PG Program — Data Science Engineering

Great Lakes Institute of Management

Python · Machine Learning · Statistics · Data Analysis · NLP · Visualization · Deep Learning · Feature Engineering.

2018

Deep Learning Specialization

DeepLearning.AI · Coursera

Improving DNNs · Structuring ML Projects · CNNs · Neural Networks · Sequence Models.

2018

Python Data Structures

University of Michigan · Coursera

Data Structures · Object-Oriented Programming.

Cert

Spark & Python for Big Data with PySpark

Coursera

Cert

Programming for Everybody

University of Michigan · Coursera

2014 — 2018

B.E. — Electrical & Electronics Engineering

Sir M. Visvesvaraya Institute of Technology

Electrical Machine Design · Engineering Mathematics · Analog Circuits · Microcontrollers · Power Electronics · Control Systems · Network Analysis.

Experience

2025.01 → Now

Technical Lead — GenAI

Elastiq AI · India

Lead the design and development of an enterprise RAG platform — multi-source ingestion, hybrid search, cross-encoder re-ranking, and document-level access control synchronized across cloud providers.

Architecting scalable distributed-database solutions via Citus DB research and implementation.
Designing distributed database architectures for enterprise-scale applications.
Mentoring cross-functional team members on robust data pipelines for large-scale ML projects.
Guiding the team through complex LLM fine-tuning and enterprise deployment strategies.
Collaborating with senior leadership on strategic product development.
Onboarded the first enterprise client for our Text2SQL solution, establishing initial market presence.

PythonLLMsRAGOpenSearchCitus DBGCPAWSDocker

2024.09 → 2025.01

Senior Software Engineer

Elastiq AI · Bangalore

Foundational AI-engineering work on Discover — the company's flagship product — covering fine-tuning, retrieval, and agent design.

Fine-tuned LLaMA 3.1 70B via advanced QLoRA / LoRA — 85% accuracy on the target task.
Engineered custom datasets and context-aware chunking for unstructured document RAG.
Built AI agents for Text2SQL and unstructured file processing inside Discover.
Built the foundational Text2SQL framework from scratch.
Integrated multiple Graph DB backends offering Classic RAG, Cypher-based RAG, and GraphRAG.
Engineered data-processing systems capable of large-scale enterprise workloads.

PythonLLaMA 3.1QLoRALangChainGraphRAGText2SQLOpenSearch

2023.05 → 2024.08

Senior Data Scientist

Acuity Knowledge Partners · Bangalore

Senior Data Scientist at a research, analytics, and technology partner to the financial services sector — asset managers, IBs, PE, VC, hedge funds, and consulting firms.

Developed a versatile analysis dashboard for an investment firm enabling experiments and KPI monitoring — 87% reduction in client analysis time.
Built a web dashboard integrated with an ESG data engine for a top-tier investment firm shaping their investment strategy decisions.
Shipped an end-to-end RAG system for parsing financial reports.

PythonStreamlitSQLDatabricksMLFlowLlama 3ChromaDBOllamaLangChainPostmanDocker

2021.12 → 2023.04

Associate — Data Scientist

TheMathCompany · Bangalore

Contributed to an enterprise AI and analytics company trusted for data-driven decision-making by some of the largest organisations across industries.

Led a 5-person team designing and shipping a flexible A/B Testing Framework for a leading pharmaceutical client to measure promotion success in revenue generation.
Co-led a 7-person team on an analyse-and-flag pipeline for sales anomaly detection with seasonality — +$125K revenue for the client in Q1.
Built an LSTM-based model forecasting customer return-to-purchase windows for supply/demand balancing — R²: 0.9012.

PythonDeep LearningDockerAirflowCodx

2021.05 → 2021.12

Data Analyst

Recruitment Smart Technologies · Ahmedabad

Built a scalable solution for periodic opportunity, operational, and advanced-analytics reporting on product performance — across multiple clients.

New-client integration: 10 days → 3 hours (97% improvement).
Operational reports: 5 days → 30 mins; opportunity reports: 1 day → 30 mins (95% improvement).
Advanced analytics report: 7 days → 15 mins (99% improvement).
Automated Power BI dashboard refreshes. Star Performer of the Month.

PythonAWS S3EC2SESPower BIDocker

2019.08 → 2021.04

Data Scientist

Sumyag Data Sciences · Bangalore

Designed and built a multi-phase, large-scale data pipeline for enriching extracted data points from documents — covering NLP, data wrangling, and text mining.

Generated 500+ features per data point to create a holistic document view via NLP, data wrangling, and text mining.
Co-developed a large matrix-multiplication model on the NumPy stack ensembling 10+ unique sources for entity classification.
Built a sequential automated pipeline producing clean datasets for custom Bayesian learning models — 10+ distinct interpretations to generate word embeddings.
Shipped an end-to-end pipeline extracting insights from claims documents for automobile and healthcare enterprise clients.
Developed a geometric model for tabular-entity extraction using asymmetric graphs and inter-point distances.

PythonNLPDeep LearningBayesianNumPy

2020.12 → 2023.05

Data Science Mentor — freelance

Independent · Topmate

Mentored college students and working professionals transitioning into data science.
Guided post-grad students on thesis projects and professionals on project management.
Conducted dozens of mock interviews for analytics and ML roles.

2016 — 2019

Earlier roles

OpenGenus · GXWeb · sabziadda · SMVIT

Pre-data-science chapter.

Content Writer (Technical) — OpenGenus Foundation, Jan – Apr 2019.
Web Developer Intern — GXWeb, Sep – Dec 2017.
Web Developer Intern — sabziadda.com (e-commerce startup), Jan – May 2017.
Student Placement Coordinator — SMVIT, Apr 2016 – Sep 2017.

06 / Services

How I can help.

Six engagements I take on most often. Pick one — or bring me a problem and we'll shape the right one.

Enterprise RAG

Architect end-to-end RAG platforms — multi-source ingestion (Drive, GCS, S3, Azure Blob, Slack), hybrid search (BM25 + kNN), cross-encoder re-ranking, and document-level ACL sync.

Agentic AI

Design and ship AI agents for Text2SQL, document processing, and custom enterprise workflows — including evaluating emerging protocols like MCP for connectors.

LLM Fine-tuning

Domain-specific tuning with QLoRA / LoRA / SFT / RLHF on open-source models (e.g. LLaMA 3.1 70B), plus custom datasets and context-aware chunking.

ML Engineering & Microservices

Build ML / DL microservices — host on GCP / AWS, instrumented with MLOps, observability, and the unglamorous infrastructure that makes the magic dependable.

Analytics & Visualisation

Develop analytic platforms and intuitive dashboards — Streamlit, Power BI, Tableau — for every flavour of analysis a problem requires.

Mentorship

I guide freshers and working professionals transitioning into the data & AI world — projects, interviews, career paths.

07 / Selected Work

Things I've shipped.

Read ↗

08 / Learn AI

Modern AI, visualized.

Five live chapters. Every panel is real math, linked. Click anything.

Foundations

Production / Serving

Operations & Trust

01 · Inside a transformer.

Every modern LLM — GPT-4, Claude, Llama — is a stack of attention layers. Pick a sentence, switch heads, click any token, and watch the math light up across all five panels in sync.

Sentence

Head

Temperature 0.8

Tokens — click one to focus its attention as the query

Attention matrix

Row = query · Col = key · Brighter = "this token looks at that token more"

What you're looking at

This updates as you interact.

The transformer takes your sentence, converts each token to a vector, then in every layer each token "attends to" every other token. The matrix on the left is that attention — for a real model, this happens 96 times in parallel (multi-head), across 80+ layers.

Switch Head above — each head learns a different pattern (position, syntax, coreference). Click a token to spotlight its row.

Next-token probabilities

Sampled at temperature 0.80 — try the slider.

Token embeddings

Each row is one token's vector. Real models use 768–12,288 dims; we show 16 for legibility.

C₁

Q · queries

"what am I looking for?"

C₂

K · keys

"what do I offer?"

C₃

V · values

"what I contribute"

Tokenize — split text into sub-word IDs.

Embed — each ID becomes a learned vector (panel B).

Project — multiply by W_Q, W_K, W_V (panels C₁/C₂/C₃).

Score — softmax(QKᵀ / √d) → attention (panel A).

Mix & predict — weighted sum of V, MLP, then logits → probabilities (panel D).

02 · The semantic map.

Embedding models turn anything — words, sentences, images, code — into points in a high-dimensional space where distance = meaning. Below is a 2-D projection of 32 concepts. Click pairs, do vector arithmetic, see clusters emerge.

Semantic space — 2-D projection

Hover any point · click two for cosine similarity · try the analogy presets below.

Cosine similarity

Click two points (or pick below) — distance measured by angle, not raw position.

↔

cosine sim

0.00

euclidean

0.00

Pick two concepts to compare.

Vector arithmetic

The famous king − man + woman ≈ queen — embeddings encode relations as directions.

Pick an analogy — watch the path on the map.

How embedding models are trained

Contrastive learning · pull similar pairs together, push unrelated apart.

Drag the slider — or hit play — watch the encoder learn:

step 0 / 1000

Anchor: "a small striped feline"

Positive (pull closer): "tabby kitten"

Negative (push away): "industrial logistics"

Loss: 2.41 — InfoNCE / triplet · same idea behind CLIP, SBERT, OpenAI ada, Cohere, etc.

Picking an embedding model

Dimensions ≠ quality. Context length, domain, and inference cost matter more for most apps.

Model	Dim	Ctx	MTEB	Best for
OpenAI `text-embedding-3-large`	3072	8k	64.6	General · multilingual
Cohere `embed-v3`	1024	512	64.5	Search · compressed
Voyage `voyage-3-large`	1024	32k	62.5	Long docs · code
`BGE-M3` (open)	1024	8k	59.6	Self-host · dense+sparse
`nomic-embed-text-v2`	768	2k	57.3	On-device · CPU
`jina-embeddings-v3`	1024	8k	59.5	Multilingual · long

03 · Fine-tuning, the cheap way.

A 70B model has billions of weights. Fine-tuning all of them needs a $50k GPU rack. LoRA trains tiny rank-r adapters instead — same effect, <1% of the parameters, single-GPU friendly. Below is the actual math, animated.

Method

LoRA rank (r) · 8

Training progress · 0%

LoRA · weight decomposition

Freeze W₀ · learn small B·A where r ≪ d. Only B·A is trained.

Training loss

Cross-entropy ↓ as the adapter learns your domain.

Train: — Val: — Step: 0

Before vs. after — same prompt

Domain: financial-report summarization. Move the training slider.

PROMPT

Summarize Acme's Q3 retail performance in our internal format.

BASE MODEL

A company reported earnings. Sales were up. Some products performed better than others.

FINE-TUNED

A company reported earnings. Sales were up. Some products performed better than others.

Pick your weapon

Cost & quality trade-offs for a 70B base.

Method	Trains	VRAM	Quality
Full FT	100% (70B)	~1.4 TB	★★★★★
LoRA (r=16)	~0.3%	~80 GB	★★★★
QLoRA (4-bit + LoRA)	~0.3%	~24 GB	★★★★
Prompt-tune	<0.01%	~16 GB	★★
Prompt eng. only	0%	0 GB	★ – ★★★

Rule of thumb: start with prompting → RAG → LoRA → full FT only if you've genuinely outgrown the rest.

04 · Retrieval-Augmented Generation.

LLMs hallucinate when asked about private data, fresh info, or anything outside training. RAG fixes it: embed your docs, look up the most relevant ones at query-time, stuff them into the prompt, and force the model to cite. Pick a query — watch the whole pipeline run.

The pipeline

5 stages. Each lights up as the query flows through.

1 · Embed query

—

→

2 · Vector search

cosine sim · top-k in index

→

3 · Re-rank

cross-encoder · refines order

→

4 · Inject

top-k → prompt context

→

5 · Generate

LLM · grounded + cited

Retrieval space — where your query lands

Each dot is one document chunk. Closer to the query (★) = more relevant.

Retrieved chunks

Top-k after re-ranking. Score = cross-encoder relevance, not just cosine.

Grounded answer

LLM sees only retrieved context — must cite by chunk ID.

Pick a query above to start.

Production RAG adds: hybrid search (BM25 + dense), ACL filters per user, chunk-aware splitting, recency boost, and answer-grounding evals.

05 · Mixture of Experts.

GPT-4, Mixtral, DeepSeek, Grok — most frontier models are sparse. Instead of activating all 400B parameters per token, a tiny router picks just 2 of 8 (or 16) "experts" to use. Same quality, ~4× cheaper to run. Click tokens — watch the router decide.

Prompt

Experts 8

Top-k routed · 2

Router → experts

Each token is scored against every expert · top-k win.

Expert specialization

Emergent — experts drift toward themes the router prefers.

Compute saved

Per-token FLOPs · live.

Dense vs. sparse — same intelligence, different cost

Why every frontier lab moved to MoE in 2024.

Model	Total params	Active / token	Style
Llama 3 70B	70B	70B	Dense
Mixtral 8×7B	47B	~13B	MoE · 2 of 8
Mixtral 8×22B	141B	~39B	MoE · 2 of 8
DeepSeek V3	671B	37B	MoE · fine-grained
GPT-4 (reported)	~1.8T	~280B	MoE · 2 of 16

The trick: serve a 671B-param model with the inference cost of a 37B dense model — that's why MoE won.

06 · KV cache at scale.

Every generated token re-uses the keys & values of every prior token. Naive caching: one GPU serves two users. PagedAttention (vLLM) treats KV memory like virtual memory — 16-token blocks, allocated on demand, shared across requests with the same prefix. The result: 5–10× throughput on the same hardware.

Strategy

Concurrent users · 8

Avg context (tokens) · 2048

GPU memory · 80 GB A100

Each cell = one 16-token page. Color = request owner. Grey = free. Striped = wasted (fragmentation / over-allocation).

Memory budget

Live · per-strategy.

Throughput

Tokens/sec the GPU can serve.

Survival tactics at scale

What real serving stacks do once they hit the wall.

Paged KV (vLLM, TGI) — 16-token blocks, no over-allocation, <4% fragmentation vs. 60-80% naive.
Prefix caching — share system-prompt KV across all users on a node. A 4k system prompt × 1000 users = saved once.
Continuous batching — new requests slot into a running batch on token boundaries instead of waiting for the slowest to finish.
KV offload & quantization — INT8 KV halves memory; offload cold prefixes to CPU/NVMe; recompute on miss.
Speculative re-use — assistant tool calls often share prefixes (system + tools schema) — cache aggressively, expire on schema change.

Quick math

Per-token KV bytes for a 70B model.

Formula: 2 (K,V) × n_layers × n_heads × d_head × bytes. For Llama-3 70B (80 layers · 8 kv-heads · 128 dim · FP16): ~320 KB/token. A 32k context = 10 MB just in KV. 100 users at 32k = 1 TB.

07 · Speculative decoding vs quantization.

Two unrelated ways to make a slow model fast. Spec decode: a tiny "draft" model proposes K tokens, the big model verifies them in parallel (one forward pass for K tokens). Quantization: shrink weights from 16-bit to 8/4-bit — model gets smaller and arithmetic gets faster, with controlled quality loss. They compose.

Draft model speed (× faster than target) · 7×

Speculation length K · 5

Quantization · FP16

Speculative decoding · token trace

Draft proposes K tokens · target verifies in one pass · accepted tokens are kept, rejected ones cause a rollback.

Weight distribution · quantization

Same layer · same weights · fewer distinct values. INT4 has only 16 levels; the model snaps to the nearest one.

Throughput vs. quality

Both techniques stacked. Live.

When to use which

Decision shortcuts.

Spec decode — best for predictable outputs (code, structured JSON, common chat). Wasted draft tokens kill the gain on creative text.
INT8 — almost free quality drop (≤1% on most evals). Default for production serving.
INT4 / GPTQ / AWQ — 2× memory savings, ~2-4% eval drop. Worth it when you need to fit a 70B on a single GPU.
Compose — INT4 + spec decode → 4-7× faster end-to-end on the same hardware, with quality within ~3% of FP16.

08 · Did the RAG actually work?

RAG looks fine until you measure it. Four canonical metrics catch the failures: context recall (did we retrieve the right docs?), context precision (was the order right?), faithfulness (did the answer stay grounded?), and answer relevance (does the answer match the question?). Compare three configs side-by-side.

Test case

Compare config

Question & ground truth

Test question, the docs that should have been retrieved, and the gold answer.

—

Should-retrieve (ground truth IDs)

—

Gold answer

—

What the system did

Retrieved chunks + generated answer. Sentences are tagged by which chunk they're grounded in (or hallucinated).

Retrieved baseline

—

Answer

—

Metrics · this run

Higher = better · 0.0 to 1.0 · RAGAS-style.

—

Config comparison

Same question · three retrieval setups.

Config	Recall	Precision	Faith.	Rel.

Build a small golden set (50-200 Q/A pairs), run it on every retriever change, watch each metric independently. A higher faithfulness with lower relevance often means you over-grounded into the wrong chunks.

09 · Where the money goes.

A polished LLM product can burn $10k/day before anyone notices. The leaks aren't dramatic — they're boring: a 4k system prompt re-sent 200k times a day, conversation history that never gets trimmed, RAG context with 30 chunks where 4 would do, retry loops on schema errors. Audit your token bill like cloud spend.

Daily requests · 200,000

Model

Mitigations applied

Where each token goes · per request

Stacked bar — system prompt, tool schema, conversation history, RAG context, user message, response. Hover for $/day attribution.

Top leaks · ranked

Auto-flagged patterns sorted by $ wasted.

The leak checklist

Print this. Pin it.

Cache system prompts — Anthropic / OpenAI prompt caching cuts repeated tokens 90%. Free 10× ROI.
Trim history — keep last K turns, summarize the rest. Cap to a hard token budget per conversation.
Right-size RAG — 4-8 well-ranked chunks beat 30 mediocre ones. Re-ranker is cheaper than long context.
Stop streaming on schema fail — early-stop when structured output is malformed; don't burn 4k tokens on bad JSON.
Per-endpoint $ caps — alert when an endpoint's $/req drifts >30%. Catches prompt-bloat regressions.
Log every call's token breakdown — input by source (system / history / RAG / user) + output. Without this you can't find leaks.

Real-world hidden leaks

War stories.

The retry loop. Structured-output failure triggers an automatic retry with the bad response appended for "context." 4 retries × ~2k tokens each = 8k wasted per failure. At 0.5% failure rate × 1M req/day = $1,200/day on a $0.003/1k model.

The runaway agent. An agent calls a search tool, gets 200KB of HTML back, stuffs it into context, calls again. Each turn doubles. One bug = 6 figures.

The dev who left logging on. Production accidentally calling verbose=True appends the full reasoning trace to every response. 3× output tokens. $8k/mo until someone checked.

The shared system prompt. One service builds the prompt fresh each call, defeating the upstream cache. 70% cache hit → 0%. $22k surprise.

10 · Agents — and the loops they fall into.

Agents are LLMs that call tools and decide what to do next. Without guardrails they will: loop on the same tool, ignore failure modes, exhaust budgets, and confidently take destructive actions. Step through a trace below — watch the loop detector, budget meter, and validators kick in.

Scenario

Guardrails

Execution trace

Each row = one tool call. Loops, schema fails, and budget breaches highlight automatically.

Budgets

Hard limits. Halt on breach.

Guardrail status

Which checks fired so far.

Loop detection — the actual patterns

What "stuck" looks like in production.

Same-call repeat — identical tool(args) twice in a row. Trivially stuck. Halt immediately.
2-cycle — alternating A → B → A → B with no progress (e.g., re-search after re-search). Detect via hash of last N actions.
Argument drift — same tool, slightly tweaked args. Often a retry loop in disguise. Halt after N similar calls (Levenshtein on args < 5%).
State stagnation — observation entropy collapses. The agent is no longer learning from results.
Token / time / cost budget — never trust the model to self-regulate. Hard caps at the framework level.

Guardrail stack — defense in depth

No single check is enough.

Input filter — PII / prompt-injection / unsafe-content classifier before the model.
Schema validation — JSON schema / regex / type check on every tool argument. Reject & resample.
Allowlist tools — minimum viable set per task. No shell.exec unless the task genuinely needs it.
Output filter — content moderation + structured-output validator on every response.
Human-in-loop — irreversible actions (delete, send, pay) require an explicit confirm step.
Audit log — every tool call, args, result, and decision. Replay any failure.

09 / Contact

Let's build something.

For projects, mentorship, or a chat about an interesting problem.

Bring a problem.
I'll help you find the shape of it.

priyanshu1996@hotmail.com

Email priyanshu1996@hotmail.com
Phone +91 7349575896
Location Bangalore, Karnataka — IN
Booking Topmate ↗
Status ● Open to collaboration