Tech Debt

An engineering-honest view of where my personal projects stand — what’s deliberately cut, what’s accumulating debt, and what each project could become with more investment.

This is a living document. It’s updated as projects evolve.

---

Why publish this?

Because “my projects are production-ready” is rarely true, and I’d rather be honest about the gap between what’s built and what’s production-hardened. Shipping a working proof-of-concept is different from shipping a product. The debt list below is how I track that gap.

---

Open debt by project

Smart Irrigation IoT

Status: Working end-to-end · ships to hardware

• Firebase API key is hardcoded in firmware — needs build-time env injection (build_flags in platformio.ini)
• Single-user Firebase structure — one device per account; multi-field deployment requires schema redesign
• SVM model retrained manually — no pipeline to retrain on new field data without code changes

Next: TFLite micro model to replace SVM (runs on-device, no cloud round-trip for predictions); MQTT instead of Firebase for lower latency and offline tolerance.

---

TrueShot (Image Validator API)

Status: Prototype — auth commented out, no live endpoint

• JWT auth middleware is wired but commented out in server.js — no endpoint is actually protected
• JWT_SECRET falls back to a hardcoded string if env var is missing — should crash on missing config instead
• No frontend — the API works but there’s nothing to demo without Postman

Next: Wire auth middleware, build a minimal demo UI (upload image → validation result), containerize with Docker.

---

EvalSheet (Handwritten Answer Checker)

Status: Prototype — OCR pipeline works, not production-hardened

• Running debug=True — needs Gunicorn + proper WSGI for any non-local use
• Similarity threshold for grading is a hardcoded heuristic — needs validation against real exam data
• No test coverage for the OCR → grading pipeline

Next: Swap Google Vision dependency for a self-hosted OCR option (Tesseract or PaddleOCR); benchmark grading accuracy on a real answer-sheet dataset.

---

ENTHRAL.IN Product Data Change Detector

Status: Working — ran in production, Windows-only

• Uses os.startfile() for opening files — Windows-only, breaks on Linux/macOS
• Pricing formula (12% + 100) hardcoded — changing supplier margin requires code edit
• All file paths hardcoded — no config file

Next: Config-driven pricing rules via YAML; cross-platform file handling; supplier-abstraction layer to support multiple data sources.

---

Blockchain NGO Donation Tracker

Status: Finished · needs testnet migration

• Deployed on Goerli testnet — Goerli is deprecated, needs migration to Sepolia
• No multi-sig protection on payout transactions — a single compromised key can drain funds
• No frontend UX for reviewing donation history beyond raw transaction hashes

Next: Sepolia migration, multi-sig payout flow (Gnosis Safe), donation history UI.

---

OMR Sheet Reading Automation

Status: Working · core algorithm is vendored

• Core OMR detection algorithm is from an open-source library — README doesn’t document this attribution
• No unit tests for edge cases (torn sheets, partial fills, multiple marks)
• Hardcoded sheet template — doesn’t generalize to different question counts or answer formats

Next: Write original algorithm or clearly document vendored attribution; parameterize sheet template via config.

---

image_hotspot Flutter Package

Status: Finished · published on pub.dev

• No Semantics wrapping — hotspot widgets are not accessible to screen readers
• No animation support for hotspot appearance/disappearance
• Test coverage is minimal — most edge cases untested

Next: Add Semantics wrapper, animation parameter, increase test coverage for pub points.

---

NO-CODE ML Environment

Status: Working · proof-of-concept scope

• Running debug=True — same issue as EvalSheet
• No dataset validation before model training — bad inputs silently produce bad models
• No model versioning — retraining overwrites the previous model

Next: Production WSGI, input validation, model registry with version tracking.

---

NMOLD Inventory Redevelopment

Status: Prototype · not end-to-end yet

• Image recognition hook exists but isn’t fully integrated — manual entry still required
• No error handling for failed Express → MongoDB writes
• Flutter app and Express backend not containerized — deployment is manual

Next: Complete image recognition pipeline, containerize with Docker Compose, deploy backend on Railway free tier.

---

Crickiaa (Gully Cricket App)

Status: Scaffold · UI only

• No business logic — scoring, match state, and leaderboard are all missing
• UI was built against a design but not connected to any data layer

Next: Implement match state machine (innings, overs, wickets), add Firebase for live score sync, basic leaderboard.

---

Skill Matrix LMS

Status: Scaffold · pipeline not working end-to-end

• Audio extraction → TTS → dub loop doesn’t produce usable output yet
• No subtitle export format (SRT/VTT)
• Tight coupling to one specific video format/language pair

Next: Wire the full pipeline using Whisper (local, free) + Coqui TTS; add SRT export; test with a real educational video.

---

Advancement opportunities (2026 framing)

Projects I built 1–3 years ago that could be significantly leveled up with current tools:

Project	Original approach	2026 upgrade path
Smart Irrigation IoT	SVM on cloud	TFLite Micro on-device, MQTT transport
AI Financial Chatbot	TF-based NLP	RAG pipeline (LangChain + embeddings)
EvalSheet	OCR + NLTK similarity	Vision-language model (GPT-4V or Gemini)
CSXIIP School Admin	Tkinter + raw MySQL	Refine.dev or AppSmith on top of a REST API
OMR Automation	OpenCV heuristics	Roboflow for training, Vision AI for detection
Skill Matrix LMS	Custom pipeline	Whisper + Coqui TTS (fully local, no API cost)

---

Last updated: 2026-06-30. See the project portfolio for write-ups on the completed work.