Sachin Kale
AI Software Engineer

Production LLM Systems,
Shipped End-to-End.

I build the whole path — fine-tuning and RAG pipelines through inference APIs, AWS deployment, and live evaluation. Three systems in production across financial news and genomics.

3
Live Systems
98K
Training Pairs
94
Automated Tests
2
Domains
What I Build

Model Adaptation

LoRA / PEFT fine-tuning, dataset curation, mixed-precision training on A100

Retrieval & RAG

Hybrid vector + BM25 fusion, cross-encoder reranking, HyDE, semantic caching

Inference & APIs

FastAPI async services, GPU endpoints, auth, rate limiting, OpenAPI contracts

Cloud & Delivery

AWS EKS and serverless Lambda, Docker, CloudFront, GitHub Actions CI/CD

Reliability

Circuit breakers, exponential backoff, multi-provider failover, graceful degradation

Evaluation

Precision@K, MRR, NDCG, LLM-as-judge faithfulness scoring, CI quality gates

1

Production RAG Pipeline

Financial News

Retrieval-augmented chat over 138 market news articles across 7 tickers, answering natural-language questions with AI summaries grounded in cited sources. Built to production standards: every answer traces back to a source, is measured for faithfulness, and is served behind a circuit breaker that degrades gracefully instead of erroring.

Retrieval Pipeline
  • Hybrid Search: Dense vector + BM25 keyword retrieval fused with Reciprocal Rank Fusion — embeddings alone miss exact tickers and financial terms
  • Reranking: Cross-encoder (ms-marco-MiniLM-L6-v2) re-scores top candidates with full query-document attention
  • Query Understanding: Intent router, entity extractor, and HyDE expansion pick a retrieval strategy per query
Resilience & Caching
  • Circuit Breaker: Opens after 3 consecutive LLM failures, auto-recovers after 60s — with a live simulation endpoint to demo it
  • Failover: Multi-provider orchestrator ranks providers by 20-minute rolling failure rate
  • Two-Tier Cache: SHA256 exact match, falling through to embedding-similarity semantic match
Evaluation & Delivery
  • RAG Metrics: Precision@K, MRR, and NDCG plus LLM-as-judge faithfulness and hallucination scoring
  • Testing: 48 backend tests (unit / integration / eval) and 46 Jest suites
  • Serverless: Lambda container + API Gateway, S3 + CloudFront frontend, GitHub Actions CI/CD
LangChainChromaDBBM25 + RRFCross-EncoderFastAPIReact 19Gemini 2.5 FlashAWS Lambda
2

Domain-Adapted LLM Platform

Genomics

An end-to-end vertical slice of LLM productization: fine-tune a base model on a curated domain corpus, serve it behind a hardened inference API, and ship a streaming chat client — deployed on Kubernetes with automated rollout. The corpus here is genetics; the pipeline is domain-agnostic.

Model — Fine-Tuning
  • Base Model: Qwen2-1.5B with LoRA fine-tuning (rank=16, alpha=32) for parameter-efficient training without catastrophic forgetting
  • Dataset: Curated 98K Q&A pairs built and validated from domain literature
  • Infrastructure: 3 epochs on NVIDIA A100 GPU (40GB VRAM) with mixed-precision (bf16)
Serving — Inference API
  • Framework: FastAPI with async request handling and automatic OpenAPI documentation
  • Backend: HuggingFace Inference Endpoints with GPU acceleration (NVIDIA T4)
  • Security: API key-based authentication with rate limiting and request validation
Client & Infrastructure
  • Frontend: React 18 + TypeScript on Vite — streaming responses, conversation history, mobile-responsive
  • Compute: AWS EKS (Kubernetes) 2-node managed cluster, ALB with SSL/TLS termination via ACM
  • CI/CD: GitHub Actions pipeline with automated ECR push and kubectl rollout
Qwen2-1.5BLoRA/PEFTPyTorchHuggingFaceFastAPIReact 18DockerAWS EKS
3

Computer Vision Annotation Platform

Any Imagery

Image annotation tool for creating high-quality labeled datasets for machine learning models — polygon annotations, time-lapse comparison with auto-detection, and multiple export formats. Domain-independent by design; the bundled demo set happens to be histopathology slides.

  • Generate labeled datasets for Vision Transformers, Diffusion Models, and Object Detection
  • Annotate individual images with regions and points
  • Add captions and metadata for training
  • Export in COCO JSON, YOLO, HuggingFace JSONL, and CSV formats
  • Import metadata files to restore annotations for any image
Annotation Features
  • Tools: Polygon regions, point markers, measurements with scale calibration
  • Metadata: Labels, notes, confidence scores, captions, and tags
  • Model-Assisted Labeling: Companion Cellpose inference API auto-detects regions in microscopy imagery (nuclei and cyto3 models)
  • Export: COCO JSON, YOLO, HuggingFace JSONL, CSV formats
Time-Lapse Comparison
  • Auto-Detection: Region tracking across image series using color profiling
  • Analytics: Area progression charts with growth statistics
  • Demo Data: Histopathology samples (Lung, Liver, Breast, Cancer)
Infrastructure
  • Frontend: React 19 + TypeScript + Vite + Canvas API
  • Backend: AWS SAM (Lambda + API Gateway + DynamoDB)
  • Deployment: S3 + CloudFront with GitHub Actions CI/CD
React 19TypeScriptCanvas APIAWS SAMLambdaDynamoDBS3CloudFront