Balaji G
build with bala
Summary
Balaji G — technical product leader and hands-on builder. Founding CTO of InzightED and, today, technical product manager at Ahoum, steering its AI product suite. I take products from architecture to roadmap — building the hard parts and leading what ships.
Experience
- Lead the Ahoum suite from roadmap to release — Talk-to-Expert, SBF and Expert chat
- Talk-to-Expert: 316 LLM personas with live Razorpay monetisation
- Designed 100→10k concurrent WebSocket scaling (Aurora Serverless v2, Fargate)
- Hierarchical-RAG evaluator + Neo4j concept-gap graph + LLM-as-evaluator
- +45% insight accuracy, OCR 94.2%→98.0%, −30% system load
- Served 10 clients
- Scalable REST APIs (Node/Express/MongoDB)
- −35% fetch latency via schema optimisation
Education
Projects
A monetised AI companion + expert platform
Technical product manager — product, LLM, infra
- 316 LLM personas Expert personas
- 70+ assessment types Assessment engine
- live Razorpay revenue Monetisation
- 10k concurrent WebSocket events Scale design
- Steer the Ahoum product suite — Talk-to-Expert, SBF and Expert chat — from roadmap to release
- Persona-prompt architecture with locked safety rails across 316 experts
- Razorpay monetisation end-to-end with HMAC-verified payments and credit passes
Next.js · Django · Capacitor · Centrifugo · Groq · Razorpay · AWS
AI that explains why a student got the marks
Founding CTO — product & architecture
- +45% (H-RAG) Insight accuracy
- 94.2% → 98.0% OCR
- −30% System load
- 10 Clients
- Owned technical + product direction as founding CTO — architecture, roadmap and hiring bar
- Architected a hierarchical-RAG NEET evaluator that goes beyond marks to concept gaps
- Neo4j concept-gap knowledge graph + LLM-as-evaluator pipeline; OCR accuracy to 98%
Django · Celery · Neo4j · PostgreSQL · H-RAG · Mistral OCR
Benchmarking when models should say "I don't know"
Research, design, scorer
- 96 real multi-session Conversations
- 4 (answerable/stale/contradiction/false-premise) Splits
- Confidently-Wrong Rate Signal
- Open source License
- Frozen pre-registration committed before any system is scored
- Confidently-Wrong Rate isolates the failures that hurt in production
- Four task splits across 96 real multi-session conversations
Python · NumPy · pytest · HF Datasets · LongMemEval
Backend that carried 90k+ monthly interactions
Backend Developer Intern
- 90,000+ Monthly interactions
- −35% Fetch latency
- +50% UX efficiency
- Built scalable REST APIs serving 90k+ monthly interactions
- Optimised MongoDB schemas for −35% fetch latency
- Integrated a React dashboard for +50% UX efficiency
Node.js · Express · MongoDB · Mongoose · React
Edge CV that protects crops, humanely
CV capstone — model + edge
- 92.4% @ 60 FPS mAP
- 1500+ images Training set
- EAI IoT Journal Published
- YOLOv5 detector deployed to Jetson Nano via ONNX at 60 FPS
- Ultrasonic deterrence validated >90%
- Published in EAI IoT Journal (DOI 10.4108/eetiot.4809)
YOLOv5 · Jetson Nano · ONNX · Ultrasonic
Software that runs an IIM, built from a wheelchair
IT Committee — lead developer
- 2000+ students Reach
- electives + election + sports portals Systems
- ecap.iimbg.ac.in Deployed
- eCAP elective Course Allocation Portal deployed at ecap.iimbg.ac.in
- SSO + RBAC encrypted election system
- Sports + admin portals with live stats and audit tracking
React · Django · PostgreSQL · SSO/RBAC
Logistics intelligence cockpit
Architecture, backend, AI gateway
- 4 (fleet · warehouse · freight · disruption) Live data domains
- cost-gated Gemini, structured output AI gateway
- Redis pub/sub + SSE Realtime
- Module-agnostic real-time pipeline over Redis pub/sub + Django Channels SSE
- Pluggable app modules wired through a central bus — no cross-module imports
- Cost-aware AI gateway with structured Gemini output and budget gating
Django 5 · React 18 · Celery · Redis · PostgreSQL · Gemini
Raw resources to finished decks
Full build — Tauri / Svelte / FastAPI
- 6 (PDF/DOCX/slides/img/YouTube/web) Ingest formats
- 12 questions Interview
- 7 sections, self-improving Prompt engine
- Multi-format ingestion with automatic research gap-filling
- Interactive 12-question interview with learned defaults
- Seven-section prompt engine with visual directives and quality rules
Svelte 5 · Tauri 2 · Rust · FastAPI · SQLite · NotebookLM
On-device form coach, gamified
Full build — Kotlin / Compose
- 100% offline, 0 network perms On-device
- 5 with form state machines Exercises
- ML Kit pose, real-time rep counting Engine
- On-device pose detection with zero network permission
- Debounced form rules + rep-counting state machines per exercise
- XP ranks (E→S), quests, achievements and a muscle-balance radar
Kotlin · Jetpack Compose · ML Kit · CameraX · Hilt · Room
Reels into a searchable second brain
Full build — Python / RAG
- fully local vector store Pipeline
- provider-swappable, graceful degrade RAG
- searchable Obsidian vault Output
- Hands-off reel ingestion via private-API polling
- Composable RAG: transcription, vision, embedding all provider-swappable
- LLM auto-tagging into Obsidian markdown with YAML frontmatter
FastAPI · SQLite · yt-dlp · Whisper · Ollama · Chrome MV3
Your phone as a camera FTP dock
Full build — Kotlin / Compose
- Wi-Fi/USB-C/hotspot live tether Transport
- 3 (Drive/Dropbox/Lightroom) Upload targets
- embedded Apache MINA FTP Server
- Embedded MINA FTP server in a session-aware foreground service
- Live gallery grid with swipe review (star / reject / skip)
- Multi-cloud auto-upload: Drive, Dropbox, Lightroom
Kotlin 2.0 · Jetpack Compose · Apache MINA · WorkManager · Room