Yash-Chindam

👋 About Me

class YashChindam:
    def __init__(self):
        self.name        = "Yash Chindam"
        self.role        = "AI / ML Engineer"
        self.focus       = ["LLMs", "RAG Systems", "Computer Vision", "NLP", "Deep Learning"]
        self.passion     = "Building AI systems that solve real-world problems"
        self.currently   = "Exploring Generative AI & multimodal research"
        self.hf_profile  = "https://huggingface.co/yashchindam"

    def say_hi(self):
        print("Thanks for dropping by! Let's build something intelligent together 🚀")

me = YashChindam()
me.say_hi()

🌍 Open Source

6 upstream projects · 5 PRs merged · 7 PRs in review · 2 packages published to a public catalog

mlflow/mlflow — the open-source AI engineering platform (~22k ★)

Merged · PR #25556 merged into master

LLM-as-a-judge scoring was unusable on Vertex AI's Claude models: every call failed with anthropic_version: Field required. The gateway's adapter_class path bypassed the provider's own _prepare_payload(), so the Vertex-specific request fields were never applied. I diagnosed it, filed #25543, and fixed it at the adapter layer with a regression test covering the judge path.

In review · a 4-PR series implementing FR #25585 — ~3,400 lines across 95 files

PR	Scope
`#25741`	Run artifact requests in the UI use the stored artifact-proxy URI
`#25742`	Logged-model artifact requests
`#25743`	Model-version artifact requests
`#25748`	Trace attachment requests

Also reported #25571 — the Bedrock Titan and AI21 completions adapters silently drop top_p and top_k.

Python · TypeScript · React · MLOps · LLM gateways

github/spec-kit — GitHub's Spec-Driven Development toolkit

Core contribution · PR #4250 merged into main

Presets could not declare that they depend on an extension, so installing one without its companion left users with a workflow that silently did nothing. I added requires.extensions to the preset manifest with strict PEP 440 validation, plus an install-time check that warns — and names the exact remediation — for missing, stale, disabled, corrupted, and version-mismatched dependencies.

Merged fixes · 3 additional PRs into main

PR	What it fixed
`#4424`	The workflows reference guide had drifted from the shipped workflow on four points — version, minimum Spec Kit version, supported integrations, and the default integration. The guide told readers the default was `copilot` when it was actually `auto`. Reconciled the documented definition with the real one and added a test that fails if they diverge again.
`#4397`	`setup-plan` emitted a JSON key named `SPECS_DIR` that held the per-feature directory, while every sibling script used `SPECS_DIR` for the specs root — the same name meaning two different paths. Renamed it to `FEATURE_DIR` across the Bash, PowerShell, and Python ports with a parity test.
`#4396`	Template composition in `common.sh` looped forever when the resolved core content contained a literal `{CORE_TEMPLATE}`, because the loop re-tested the string it had just rewritten. Matched the already-correct PowerShell and Python semantics instead of inventing new ones.

Community catalog · published, listed, and maintained at v0.1.1

Package	What it does
`speckit-inventory`	Read-only extension that derives every live `FR-`/`NFR-`/`SC-`/`T-` ID from existing specs and returns focused per-task context packs instead of whole-file dumps
`inventory-alignment`	Preset that makes the agent classify each requirement against the live set before writing, so a reworded requirement is updated rather than duplicated

Both shipped at v0.1.0, then updated to v0.1.1 when I found the published entries pointed at a download_url that 404s — source at spec-kit-inventory-alignment.

In review · PR #4488 — a bundled github extension for taskstoissues, implementing someone else's feature request (#4421). 1,541 lines across 13 files.

Python · CLI tooling · agent workflows · zero runtime dependencies

Also contributed to

Project	Contribution
obra/superpowers	`#2263` — brainstorming keeps the open question open when the user asks for context, instead of treating the request as an answer · `#2259` — re-resolve and verify relative links when skill content moves
langchain-ai/langgraph	Reported `#8673` — checkpoint serialization rejects any int outside the 64-bit range, though `Decimal` of the same value round-trips fine
crewAIInc/crewAI	`#7042` — tools returning coroutines inside an already-running event loop
NousResearch/hermes-agent	`#89498` — the curator dropped configured fallback chains instead of forwarding them

🛠️ Tech Stack

Languages & Frameworks

AI / ML

LLMs & Models

Vector Databases & RAG

Data & Tools

🚀 Featured Projects

🛡️ Secure MCP Multi-Agent Research Platform

A governed research system where five FastMCP servers are driven through a single gateway that authorizes, meters, circuit-breaks and audits every tool call. Workflow state transitions reject invalid stage skipping, claims require supporting evidence, and reviewer approvals bind to one exact server, capability, resource and argument digest.

Security model: Rego policy-as-code that fails closed · Keycloak OIDC with an asymmetric algorithm allowlist · capability registry whose discovery reveals only what the caller may see · tenant-isolated jobs and evidence Stack: Python 3.12 FastMCP FastAPI OPA/Rego Keycloak PostgreSQL Docker

🔀 Production Local-LLM Inference & Routing Platform

An OpenAI-compatible control plane that routes each request across local model tiers on inferred task, privacy class and policy score. Every response records the selected model, its immutable revision, the candidate count and the reason it was chosen — so a routing decision is auditable after the fact.

Engineering: three separately reported CI layers (unit/static, integration, Playwright) gate the release image · CD publishes a versioned OCI artifact · production startup refuses the dev key Stack: Python FastAPI Ray Serve vLLM TypeScript Playwright Docker

🔍 Self-Optimizing Production RAG Platform

RAG that cites or abstains — never guesses. The query path is an explicit state machine (classify → clarify → rewrite → decompose → retrieve → generate → verify → repair), so an ambiguous question gets clarified and an unverified answer is narrowed to its best-supported sentence before falling back.

Self-optimizing: an evaluator scores retrieval recall, grounding and forbidden claims deterministically rather than by LLM judgment, then a control loop perturbs one pipeline field at a time within reviewer-approved bounds and keeps only Pareto-optimal candidates that never regress authorization, latency or quality Stack: Python LangGraph DSPy Qdrant OpenSearch Neo4j Presidio

🔒 LLM Security & Agent Guardrail Gateway

A policy-aware gateway that inspects untrusted prompts, retrieved context, model output and proposed tool actions before they cross a security boundary. Risky actions return an approval token bound to the exact action digest, tenant, expiry, and one-time use.

Adversarial evaluation: a PyRIT-style red-team suite covering direct and indirect injection, multi-turn jailbreak, encoded instructions, cross-tenant access, tool privilege escalation and MCP tool poisoning — scored against a committed baseline so a regression fails CI instead of landing silently Stack: Python FastAPI PyRIT Bandit Trivy Playwright Docker

🧬 Drug-Protein Interaction Prediction

Predicts drug-protein binding strength (Ki values) using OpenAI's CLIP vision encoders on 2D molecular images and protein sequence logos.

Results: RMSE 0.6041 · MSE 0.3649 Stack: PyTorch CLIP RDKit Transformers CUDA

🏦 Finlyzer — Bank Statement & Credit Report Analyzer

Financial document analysis that uses Google Gemini vision to extract, categorize and summarize transactions from bank statements and credit reports, then surfaces recurring patterns, credit score, open loans and overdue status.

Throughput: 9-step automated pipeline · parallel processing across up to 4 API keys via ThreadPoolExecutor Stack: Gemini Pro FastAPI Streamlit GCS Docker

🗂️ More Projects (click to expand)

Project	Description	Stack
Intelligent Claims Document Processing	ClaimLens AI — agentic pipeline for US commercial property insurance across 8+ document types	`LangGraph` `Azure OpenAI` `Pydantic`
AI-Powered Natural Language to SQL Engine	NaturalSQL — plain English to executable PostgreSQL via SQLCoder-7b-2	`SQLCoder-7b-2` `PyTorch` `Cloud SQL`
AI Voice Onboarding System	Modular AI onboarding framework with multi-LLM support and guided setup workflows	`LangChain` `LlamaIndex` `FastAPI`
RFP Document Info Extraction via LLMs	Structured extraction from RFP documents, 85–92% accuracy at 30–60s/doc	`GPT-4` `LangChain` `PyPDF2`
Vision-Based Entity Extraction	OCR + NER over forms, invoices, ID cards and business cards — entity F1 80–92%	`EasyOCR` `spaCy` `YOLO`
Multi-PDF Chatbot with RAG & FAISS	Chat across many PDFs at once with RAG over FAISS	`Mistral Nemo` `FAISS` `LangChain`
ECO2 — Environmental ML Platform	Carbon footprint tracking, climate modeling and biodiversity assessment	`GeoPandas` `NetCDF4` `Plotly`
YouTube Video Summarizer	Transcribes and summarizes YouTube videos with a Streamlit UI	`Gemini Pro` `Streamlit`
RAG w/ LLaMA2 + LangChain + ChromaDB	End-to-end RAG pipeline using LLaMA 2	`LLaMA 2` `ChromaDB`
PDF Chatbot with RAG	Conversational PDF Q&A with RAG architecture	`RAG` `FAISS` `LLMs`
Image Captioning	Deep learning-based automatic image captioning	`PyTorch` `CNN` `LSTM`
License Plate Recognition	Automatic license plate detection and OCR	`OpenCV` `OCR`
Text Summarization — BART	Abstractive text summarization with BART	`BART` `Transformers`
Research Paper Title Generator — BART	Fine-tuned BART for academic title generation	`BART` `Transformers`
Movie Title Generator — Flan-T5	Flan-T5 fine-tuned for cinematic title generation	`Flan-T5` `HuggingFace`
Predicting Credit Card Approvals	ML classifier for credit card approval prediction	`scikit-learn` `Pandas`
RAG Implementation & Prompt Optimization	Benchmarking and optimizing RAG prompt strategies	`RAG` `LLMs`

📊 GitHub Stats

🤗 Hugging Face

Published models & datasets on Hugging Face

Resource	Link
🧬 Drug-Protein Interaction Model	yashchindam/Drug-Protein-Interaction-Prediction-Using-CLIP-and-Deep-Learning
📦 Drug-Protein Dataset	datasets/yashchindam/Drug-Protein-Interaction-Prediction-Using-CLIP-and-Deep-Learning

🏆 Highlights

"The best way to predict the future is to invent it." — Alan Kay

Navigation Menu