01 / Applied ML Engineer

Mihir Mohite

I build retrieval and agentic ML systems — RAG pipelines, NL-to-SQL agents, and the occasional rover. Currently ML intern at BeyondBot and president of CoDeC @ MIT-WPU.

Get in touch Download résumé ↓
Mihir Mohite
Role
ML Intern · BeyondBot
Working langs
Python · C++
RAGAS Δ
+0.26 hybrid v naive
Corpus
1,210 pp · IN law

Building a legal-compliance RAG system at BeyondBot over a 1,210-page corpus of Indian industrial, manufacturing and environmental law. Benchmarked five retrieval modes on a 315-query RAGAS harness — hybrid retrieval reached 0.77 against 0.51 for the naive baseline.

mistral ocr → regex router → lightrag (kg + vector) → agno · gemini 3 flash → fastapi sse

recognitions · sih r2 ′24 · circuit heist runner-up ′24 · dataquest finalist · hackmitwpu finalist

02 / Experience

Where I've worked

2025.10 — now

Machine Learning Intern

BeyondBot Technology Pvt. Ltd. · Pune

Designed legal-compliance RAG over a 1,210pp Indian-law corpus. RAGAS harness across 5 retrieval modes on 315 queries — hybrid 0.77 vs 0.51 naive. Built a custom OpenCV visual chunker.

read →
2025.07 — 2025.10

Software Intern

Heera Software Pvt. Ltd. · Pune

Address-confidence scoring service on the Google Maps API with fuzzy matching and reverse-geocoded cosine similarity. Scored ~13,500 records and sharpened address confidence for logistics clients including Red Bull.

read →
03 / Projects

Selected work

ForeSites — Construction Site Management
pilot · 2026

ForeSites — Construction Site Management

3-person team · pilot with a Pune real-estate developer

AWS Lambda agent (Gemini) doing streaming NL-to-SQL over a 9-table Postgres schema, with per-session memory. WhatsApp ingest service (Node + S3, Supabase-backed) for snag reports via text, voice and image.

Read case study →
2025

AgroSense — Precision Agriculture (PBL4, MIT-WPU)

team

UAV photogrammetry pipeline (RGB → orthomosaic → ExG/VARI vegetation indices → K-means zoning) fused with an ESP32 ground module (Modbus soil sensor, GPS waypoint guidance, on-device ML for NPK estimation). Firebase + Vercel dashboard with a multilingual AI assistant.

Read case study →
2025

EKG — query-driven Neo4j mini-graphs

solo

Demo ingesting SQL / JSON / TXT into query-driven Neo4j mini-graphs with a Gemini NL-query interface.

Read case study →
03.5 / Earlier
04 / Leadership

CoDeC, MIT-WPU

2025.10 — now

President

CoDeC · MIT-WPU

Ran Trifecta Challenge 2026 — a three-day flagship symposium drawing 350+ participants (87 teams) from MIT-WPU and external engineering colleges across Full-Stack, ML, and Competitive Programming tracks. Led a core team of 15–20; secured GeeksforGeeks, HackerRank and AlgoZenith through direct outreach.

read →
2025.01 — now

ML Projects Division Lead

CoDeC · MIT-WPU

Drive AI/ML project initiatives and run knowledge-sharing sessions for the club.

Mihir Mohite presenting at CoDeC, MIT-WPU
presenting at codec · mit-wpu · 2025
05 / Education
2023 — 2027 (exp.)

MIT World Peace University, Pune

B.Tech Electronics & Communication Engineering · Specialization in AI & ML

CGPA 8.62 / 10.00

06 / Skills

Tools I work with

Languages
C++PythonJavaScriptSQL
CS Fundamentals
Data StructuresAlgorithmsOODComplexityProblem Solving
Retrieval / RAG
RAGGraphRAGLightRAGRAGAS-style evalSentence TransformersBGE
Vector DBs
PineconeMilvus
LLM Platforms
Vertex AIGemini APIAgno
ML / Vision
PyTorchTensorFlowKerasCNNTransformersOpenCVOCRIBM DoclingMistral OCR
Web & Cloud
React.jsNode.js / ExpressFastAPIRESTSSEAWS (Lambda · S3 · EC2)GCP (Vertex AI)serverless
Databases
PostgreSQL (Supabase)MySQLSQLiteNeo4j
Tools & Hardware
GitMATLABWhatsApp Business APIMeta APIArduinoESP32Raspberry PiLiDAR
07 / Contact

Let's talk

Open to internships, research collaborations, and conversations about retrieval systems. Email is fastest.