Skip to content

Latest commit

 

History

9 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 

Repository files navigation

Sanjay Sakhinala

AI/ML Engineer — Agentic RAG · LLM Evaluation · Reinforcement Learning

Portfolio LinkedIn Email Resume

Hyderabad, India · Open to entry-level AI/ML and software engineering roles


About

AI/ML engineer working on retrieval-augmented generation, agentic systems, LLM evaluation, and reinforcement learning. B.Tech in Electronics and Communication Engineering, 2025.

The thread running through everything below is making LLM output verifiable — citation grounding enforced at the schema level, NLI-based claim-level faithfulness checks, and offline eval harnesses that turn "the answer looks right" into a number you can regress against. Author of 3 IEEE publications and 1 Indian patent.

Currently extending IndicRAG's evaluation harness and pushing AgentOps toward a drop-in tracing SDK for third-party agent stacks.

Languages Python · SQL ML / RL PyTorch · TensorFlow · scikit-learn · LightGBM · Stable-Baselines3 · Optuna · OpenCV NLP / LLM HuggingFace Transformers · LangChain · LangGraph · CrewAI · BGE-M3 · Gemini Backend / Infra FastAPI · SQLAlchemy 2.0 · PostgreSQL · pgvector · ChromaDB · FAISS · Redis · Docker · Nginx · Prometheus · MLflow · GCP / Vertex AI · Git · Linux Data pandas · NumPy · SciPy · Matplotlib · Streamlit


Featured Work

🔷 IndicRAG — Multilingual Agentic RAG

Commits Last commit License

Problem RAG systems answer confidently in English and fail quietly in Indic languages, with no signal for when a citation doesn't actually support the claim.
Approach LangGraph StateGraph — planner → tool selector → executor → generator → reflexion evaluator — with 6 tools and dual-gate verification (bge-reranker-v2-m3 faithfulness + Gemini completeness, threshold 0.75). Hybrid BGE-M3 dense + BM25 retrieval fused with RRF (k=60), then cross-encoder and ColBERT MaxSim reranking, then NLI claim-level faithfulness verification.
Result Precision@5 1.000, Recall@5 0.917, MRR 1.000, Citation Grounding 0.938 on manually labeled relevance judgments, across 10 Indian languages and English. Reflexion loops converge within 3 iterations under a 45 s wall-clock budget.
Stack LangGraph BAAI/bge-m3 ChromaDB Gemini FastAPI PyMuPDF Prometheus Nginx Docker Compose

Solo-developed continuously since November 2025 — 100+ commits through v2.4.0, PR-based review workflow, semantic versioning, documented rollback procedures. The hardest problem wasn't retrieval quality but knowing retrieval quality: multilingual relevance judgments had to be hand-labeled before any of the numbers above meant anything. Building the reflexion loop taught me that a self-correcting agent without a hard iteration and wall-clock budget will happily spend unbounded tokens convincing itself.


🔷 Meridian UM — AI-Augmented Prior Authorization

Problem In regulated review workflows, an ungrounded model finding is worse than no finding — and prompt changes are usually evaluated anecdotally.
Approach Citation grounding enforced at the schema level (findings must resolve to a document span or policy-clause ID; uncited findings rejected, low-confidence routed to review) plus a PostgreSQL CHECK constraint blocking any denial lacking a recorded human attestation. 11-state forward-only case workflow: step-skipping rejected, repeats idempotent.
Result A property test fails if either invariant is weakened. Offline eval harness reports accuracy, per-label P/R/F1, and citation validity over ~150 labeled synthetic cases.
Stack FastAPI (async) SQLAlchemy 2.0 PostgreSQL 16 pgvector OCR JWT Apache-2.0

Synthetic data only. No real patient data anywhere in this project.


🔷 AgentOps — LLM Agent Observability & Evaluation

Problem Agent quality regressions are invisible without per-run traces and per-run metrics — you find out from users, not from CI.
Approach Async FastAPI + PostgreSQL over a 12-table trace schema exposed via REST and WebSocket, plus a Python tracing SDK that buffers steps client-side and captures exceptions as failed runs, so instrumentation never crashes the host application.
Result Per-run faithfulness, retrieval relevance, hallucination score, and unsupported-claim extraction, via a dual-path evaluation engine — zero-cost token-overlap heuristics by default, swappable to a Gemini LLM judge. Prompt Lab A/B-tests versions on success rate, cost, latency, and hallucination delta.
Stack FastAPI PostgreSQL Next.js 14 React Flow Python SDK Docker Compose

🔷 RL for Antenna Design Optimization — Accepted, WAMS 2026

Problem Each CST Studio evaluation is expensive; a genetic-algorithm baseline needed ~1,500 of them to converge on a 28 GHz patch antenna.
Approach Soft Actor-Critic trained over 2.68M steps against a blended surrogate ensemble (2×LightGBM + MLP + RidgeCV, <5 ms inference) with uncertainty-aware reward shaping and an active-learning loop that re-simulates top candidates to limit model drift.
Result 420 simulations — a 72% reduction. Outperformed GA, PSO, DQN, PPO, and DDPG: −52.2 dB return loss, 8.93% FBW, 5.272 dBi gain, 80.6% radiation efficiency. Converged on a quad-core CPU with no GPU, using curriculum learning (32 → 64 → 128 step episodes) with VecNormalize.
Stack PyTorch Stable-Baselines3 LightGBM Optuna CST Studio API

More projects
Project What it is
ai_data_scientist Seven sequential CrewAI agents take a raw dataset to cleaned data, trained models, charts, and an executive report. Every agent is bound to deterministic pandas/scikit-learn/SciPy tools, so the LLM orchestrates while the tools compute. SQLite experiment store, LiteLLM provider abstraction, exponential-backoff retry on HTTP 429.
landuse-cnn-eurosat Custom 3-layer CNN trained from scratch in PyTorch — 89% test accuracy on 27,000 EuroSAT images across 10 classes, with Grad-CAM hooks on the final Conv2d layer verifying the model learns spatial rather than texture features.
CipherChat End-to-end encrypted messenger implementing the Signal Protocol — X3DH key agreement and Double Ratchet — with AES-256-GCM, TLS transport, perfect forward secrecy, and TOFU key verification.
ResolverLab DNS benchmarking across 50+ filtered providers — block accuracy, latency, cache behavior — over UDP, TCP, DoH, and DoT, with automated HTML reporting.

Research

Publications

  • Uncertainty-Aware Reinforcement Learning System with Blended Surrogate Models for Electromagnetic Structure Optimization — Accepted, WAMS 2026, BVRIT Hyderabad.
  • Bandwidth Enhancement of Slotted Hexagonal Patch Antenna for 6G Ultra-Fast Data Transfer and Brain-Computer Interface — ICMOCE 2025, IIT Bhubaneswar. DOI: 10.1109/ICMOCE64100.2025.11076991
  • Bandwidth Optimization of Slotted Circular Patch Antenna for 6G Ultra-Fast Data Transfer and Brain-Computer Interface Application — 16th ICCCNT 2025, IIT Indore. IEEE Xplore publication pending.

PatentDesign of Hexagonal Patch Antenna at 28 GHz, Indian patent application No. 202541014595 A, published March 2025.


Training

Programs — Generative AI, SkillHive Connect (Aug–Dec 2025) · AI/ML, Google for Developers (Jan–Mar 2025) · Cybersecurity, Palo Alto Networks (2024) · Data Science, Altair (2024) Certificates — Google Data Analytics · Google Cybersecurity · NPTEL Cloud Computing · Google Cloud skill badges (TensorFlow on GCP, Vertex AI Pipelines)


Fun Facts

  • 📡 I came to machine learning through antenna design — my first optimization problem was a physical object, not a dataset.
  • 🔍 I use Grad-CAM to check whether a CNN is cheating, not to make pretty heatmaps. Texture shortcuts don't show up in an accuracy number.
  • 🌐 The handle is DNSdecoded, and there's a resolver benchmark behind it: 50+ providers across UDP, TCP, DoH, and DoT.

GitHub · LinkedIn · Portfolio · Email

"The purpose of computing is insight, not numbers." — Richard Hamming

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors