TenderLens

TenderLens is a tender-intelligence SaaS for Indian public procurement. It finds every open tender on 35 official NIC GePNIC portals (central, PSU, defence and 27 states and UTs), helps a contractor understand the tender documents in minutes with a Copilot that runs on our own open-weight LLM, and lets a team manage each bid until it is won. The build spec is docs/PLATFORM_V2.md; section 5 sets out how it competes with incumbents such as TenderDetail: published prices, AI on every plan, page citations, visible data freshness, and an API.

Find

• Search with typo tolerance, prefixes and tender-ID lookup, disjunctive facets (sector, state, PIN area, value, type), an India map and a sector view (Postgres full-text + pg_trgm; no Elasticsearch)
• Recommended for you: open tenders matching the company profile's states, sectors and turnover (value ≤ 3 × annual turnover), each with its reasons
• Email alerts with the same search semantics, sent after each crawl
• /coverage: every portal with its open tenders and last successful crawl

Understand (Copilot)

• Upload the NIT/BOQ PDFs (OCR for scanned pages). The Bid Brief extracts EMD, fees, dates, pre-bid meeting, turnover and experience criteria, penalties and required documents by rule, with page citations, in milliseconds
• Eligibility: the brief compared with the company profile (eligible / not eligible / unknown, check by check)
• Ask: hybrid retrieval (Postgres FTS + pgvector, RRF) → a self-hosted LLM (llama.cpp, Qwen3.5-4B GGUF, any OpenAI-compatible server) → a grounding check that every number and date appears in a cited passage, else the answer is refused. If the LLM is down, the answer is built from the best passages instead (mode: "extractive"). No paid LLM API.

Manage

• Workspaces with owner/admin/member roles, invitations and seats
• Bid pipeline (watching → preparing → submitted → won/lost/dropped) with a kanban board, closing-soon reminders, and an iCal feed (bid due, pre-bid meeting, bid opening) for Google Calendar or Outlook
• CSV export, OCDS 1.1 release packages, REST API keys
• Plans with published INR prices and quotas; self-serve Razorpay subscriptions
• Installable PWA, light/dark mode

Stack: Python 3.12 · Django 5.2 + DRF · PostgreSQL 16 (full-text, pg_trgm, pgvector) · Celery 5 + Redis · llama.cpp · sentence-transformers (multilingual-e5-small) · OCRmyPDF · React 19 + TypeScript + Tailwind 4 + TanStack Query · Docker · Caddy

Doc	What is in it
docs/PLATFORM_V2.md	Build spec, API contract, competitive positioning, status and next steps
docs/SOURCE_NOTES.md	The 35 portals, rejected ones, why award data is not collected, cross-portal duplicates
docs/SAAS.md	Workspaces, pipeline, calendar, recommendations, API keys, exports, billing
docs/DEPLOY.md	Production runbook: VPS, Caddy, backups, monitoring, monthly cost
LLM kit (not on GitHub)	Model choice, QLoRA fine-tuning, GGUF export, LLM evaluation. Kept on the founder's disk next to the repo (../tenderlens-llm, linked as llm/)
docs/INTERVIEW_NOTES.md	The non-obvious engineering decisions, explained

Quickstart (development)

cp .env.example .env            # then set DJANGO_DEBUG=true and DEV_LOGIN_ENABLED=true
make up                         # postgres, redis, web, workers, beat, mailpit, caddy → http://localhost:8080
make crawl SOURCE=central MODE=full

No docker compose plugin on the host? Every compose target takes COMPOSE=; run Compose from the official CLI image instead:

make up COMPOSE="docker run --rm -v /var/run/docker.sock:/var/run/docker.sock -v $PWD:$PWD -w $PWD docker:27-cli compose"

Running Django on the host instead (what the tests do): start only postgres and redis, run uv sync --frozen --extra copilot once (the Copilot's embedding model needs torch, which is an optional extra so the Vercel build stays under its 500 MB function limit), then uv run --frozen python manage.py migrate && uv run --frozen python manage.py runserver. Settings come from the process environment (not from .env), so export DJANGO_DEBUG=true DEV_LOGIN_ENABLED=true for the any-email dev sign-in, and CELERY_TASK_ALWAYS_EAGER=true if no worker is running. The Copilot's LLM is optional: start it with COMPOSE_PROFILES=llm (needs the model file, e.g. Qwen3.5-4B-Q4_K_M.gguf, in llm/models/); without it, answers are extractive. The frontend dev server (npm run dev in frontend/) proxies /api to localhost:8000. If the host's node is broken, use docker run --rm -u $(id -u):$(id -g) -e HOME=/tmp -v "$PWD/frontend":/app -w /app node:22-bookworm-slim npm ….

Numbers from a real crawl

Integration run, 6 October 2026 (all 35 portals, live, through the real pipeline; a bounded check, not a full crawl: up to 25 organisations and 15 detail pages per portal, ≤ 1 request/second per host):

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗