Ajeet Singh

Ajeet Singh

AI/ML Engineer · Bengaluru, India

I build production AI systems you can verify.

Four years designing and shipping Generative AI, RAG, multi-agent and document-intelligence systems, from architecture to deployment on AWS.

Now: document intelligence for U.S. real estate records at IDA Automation

Knowra · try it

Illustrative data

Same question, four roles. Access rules decide what the model is allowed to read.

Asked as Finance

“What did we spend on contractors last quarter?”

Q3 contractor spend was $412,800, about 6% above Q2, driven by two engineering contracts extended through September.

1 FY26 Contractor Ledger.xlsx · row 118
2 Q3 Close Summary.pdf · p.4

4 yrs

Building software, ML and Generative AI systems

10

AI engineers led as Team Lead at HeyBobo

~80%

Less manual document processing with my extraction pipeline at IDA

0 / 200+

Restricted-document leaks across Knowra's permission tests

Selected Work

Systems where every answer points back to its source.

Featured · Personal project · 2026 to Present

Knowra

Permission-aware enterprise RAG platform

Turns an organisation's documents into a searchable knowledge base that answers with cited sources, restricted strictly to documents the person asking is authorized to view. Full stack: ingestion, retrieval, chat and admin tooling.

Top 5

Right source retrieved in testing

~20%

Fewer hallucinations after re-ranking

2–5 s

Typical time to a cited answer

FastAPI React/TS PostgreSQL FAISS + BM25 Docker Compose

How a question becomes a cited answer

01 Question plain English
02 Intent router skip or search
03 Access filter fail-closed RBAC
04 Hybrid retrieval FAISS + BM25 → RRF
05 Re-rank cross-encoder
06 Cited answer page + passage

Professional work · IDA Automation

Document Intelligence

Structured data from U.S. real estate records

Pipelines that extract, validate and structure fields such as legal land descriptions from mortgage, title and ownership documents, paired with a computer-vision and vision-language pipeline that reads assessor parcel maps.

100+

Documents processed per day

~80%

Less manual processing effort

~90%

Fewer extraction errors

  • Fine-tuned LLMs: ~15–20% more accurate than a general-purpose LLM baseline
  • Classical CV + VLM to detect parcel geometry and compass orientation
  • FastAPI microservices on AWS Bedrock, SageMaker, EC2 and S3

Python · Computer vision · VLMs · RAG · AWS

Read the case study & interactive demo →

Production backend · HeyBobo · 2025

fitness.ai

LLM-driven workout planning backend

A production FastAPI backend that generates personalised standard and athlete-specific workout plans, with agentic follow-ups that remember the conversation. I owned the architecture and implementation end to end.

  • LangGraph/LangChain agent layer with conversation memory
  • Response caching reuses past generations to cut token cost and latency
  • Auth and RBAC; pluggable LLMs via Ollama or custom APIs

Python · FastAPI · LangGraph · LangChain · Ollama

Code on GitHub ↗

Experience

Where I've built.

Full résumé (PDF) ↗

Mar 2026 to Present

Bengaluru, India

AI/ML Engineer · IDA Automation

  • Design and build production Generative AI document intelligence for U.S. real estate records including mortgage, title, ownership, commercial and residential documents.
  • Built and deployed an AI extraction pipeline processing 100+ documents a day, cutting manual processing effort by ~80%.
  • Fine-tuned LLMs for structured extraction, improving accuracy ~15–20% over a general-purpose LLM baseline and reducing extraction errors by ~90%.
  • Designed a multimodal pipeline for assessor parcel maps, pairing classical computer vision with a vision-language model to detect parcel geometry and compass orientation.
  • Architected FastAPI microservices on AWS (Bedrock, SageMaker, EC2, S3) for ingestion, AI orchestration and retrieval; tuned prompts and context to reduce hallucinations.

Jul 2025 to Feb 2026

Bengaluru, India

AI/ML Engineer, Team Lead · HeyBobo

  • Led a team of 10 AI engineers building a multi-domain assistant for health and education, overseeing architecture decisions, sprint planning and mentoring.
  • Designed multi-agent workflows in LangGraph and LangChain with multi-step reasoning, tool calling, conversational memory and orchestration.
  • Fine-tuned open-source LLMs (LLaMA, Qwen, Falcon) on health and education datasets.
  • Deployed training and inference on AWS SageMaker, Bedrock and S3, served through FastAPI in Docker.

Sep 2022 to Feb 2025

Bengaluru, India

Software Engineer · GL Communications

  • Built regression models forecasting daily revenue for 50+ products from 100K+ historical sales records, producing 7-day forecasts for sales planning.
  • Improved forecast accuracy by ~15% (MAE/RMSE) over a seasonal naive baseline through feature engineering, cross-validation and tuning.
  • Wrote optimised SQL and Python ETL pipelines (Pandas, NumPy) feeding ML models and analytics dashboards.

Skills

The toolkit.

Generative AI & agents

LLMs RAG Multi-agent systems LangGraph LangChain Tool calling Fine-tuning Multimodal LLMs Guardrails

Retrieval & search

Hybrid search Dense + BM25 Embeddings FAISS Cross-encoder re-ranking Reciprocal Rank Fusion Semantic chunking

ML, deep learning & vision

PyTorch Hugging Face Sentence Transformers scikit-learn Computer vision NLP

Models & providers

AWS Bedrock OpenAI Gemini Groq LLaMA Qwen Ollama

Backend & APIs

Python FastAPI REST APIs Microservices JWT auth RBAC React TypeScript

Cloud, data & delivery

AWS SageMaker EC2 S3 Docker Docker Compose PostgreSQL SQL Git

About

AI that holds up in production.

Ajeet Singh - AI/ML Engineer

I'm an AI/ML engineer who owns systems end to end, spanning architecture, backend engineering, model fine-tuning and cloud deployment. I started in software engineering, building forecasting models and data pipelines, and moved into Generative AI, where I've led a team of ten and shipped agentic and document-intelligence systems.

The thread through my work is making AI output something you can check. At IDA Automation, that means turning scanned property records and parcel maps into structured, validated data. Knowra applies the same idea to company knowledge: every answer cites its source, and access rules are enforced before the model sees anything.

Traceable by design

Outputs point back to the page, region or passage they came from.

Production over demos

Validation, caching, auth and automated tests: the essential layers that make AI dependable.

Right tool first

Classical methods where they win; models where they earn their place.

B.E. Computer Science, Visvesvaraya Technological University (2018–2022) · English, Hindi

Contact

Let's build something useful.

Open to AI/ML engineering roles and selective consulting projects across enterprise RAG, intelligent agents, and document automation.

✓ Copied to clipboard!