02NeLabs AI Task Management System2026
Turning a full inbox into a ranked task queue
Business tasks arrive through email, WhatsApp, phone calls, calendar invites and documents. People track them in their heads, and the urgent ones are rarely the ones that shout loudest. This system extracts the tasks, ranks them, and schedules them.
01The problem
Task capture was manual and lossy. An actionable request sent by WhatsApp at 9pm had the same chance of being done as one in an email thread — which is to say, none — and there was no single place to see what was actually outstanding.
Two things had to be true for the system to be useful: extraction had to be reliable enough to trust, and ranking had to reflect business consequence rather than message order.
02What I built
- Five-channel ingestion — email, WhatsApp, calls, calendar invites and documents normalised into one event shape.
- Rule-first extraction — regex and pattern rules for deadlines, meetings, follow-ups, urgency, action items and document requests, plus natural-language date parsing (today, tomorrow, next Monday, end of week, end of month, in N days).
- Model fallback, not model default — where rule confidence drops below threshold, the message is sent to Llama 3.1 Nemotron 70B on NVIDIA NIM to extract structured task fields.
- Seven-factor priority scoring — urgency, business value, sender importance, deadline proximity, context relevance, required action and capacity impact, combined into a single weighted score.
- Deduplication — the same request arriving twice does not create two tasks.
- Planner, dashboard and goal tracking — an AI planner proposes a schedule; completing a task recalculates progress on the goals it is linked to.
03How a message becomes a ranked task
04The decisions that mattered
Rules first, model second. Extraction runs on every inbound message, so it sets the cost and latency floor for the whole system. Rules cover the common shapes deterministically; the 70B model handles the residue where rules are not confident. Running the model on everything would have been slower, more expensive and less predictable.
Seven factors, not a date sort. Sorting by deadline is easy and wrong — a task can be due tomorrow and still not matter if it is low value from a low-importance sender with no capacity behind it. The weighting makes those trade-offs explicit instead of leaving them to whoever scans the list first.
Deployment is part of the feature. The system runs on an Ubuntu server with PM2 supervision, Caddy in front, Cloudflared for the tunnel and UFW for the surface. AI features that only run on a laptop do not get used.
05Where it stands
Deployed and running on a production VPS. Every task keeps a link back to the message it came from, so a wrong extraction is traceable to its source rather than silently wrong. The natural-language date parser covers the phrasings people actually type — end of week, EOD, next Monday, in three days.
06What I would change next
- A labelled evaluation set — extraction quality should be measured as precision and recall on annotated messages, not judged from a demo.
- Confidence calibration — the rule-versus-model handoff threshold is tuned by judgement today; it should be derived from measured accuracy.
- Feedback into ranking — when a user reorders or dismisses a task, the weights should learn from it.