leadbybuild.ing

WorkSelected builds

Built after hours.

Pet projects I build in my own time, to learn something properly or to fix a problem I keep running into.

My professional work stays confidential, so none of it is here. Everything on this page is personal, built and shipped on my own.

  1. 01ToolLensFinds which MCP tools AI agents confusePython · MCP · LLM evals
  2. 02IPO RadarSelf-healing IPO tracker, with a facts APIPython · REST API · Data
  3. 03Next buildIn progress

01 · Developer tooling · 2026

ToolLens

Find out which MCP tools your agents confuse, before your users do.

What
An evaluation harness for Model Context Protocol servers
Stack
Python, official MCP SDK 2.x, httpx, rich
Models
Ollama, MLX, LM Studio, vLLM, Anthropic, OpenAI, OpenRouter, BM25
Ships as
CLI, self-contained HTML report, CI gate · MIT

01The problem I hit

My agents kept calling the wrong tool, and nothing told me.

I was wiring MCP servers into agents, mostly on small local models. They'd pick list_events when they needed find_free_slots, or call nothing at all. The only way to find out was trial and error in a chat window.

The idea

One command.

Point ToolLens at an MCP server and a set of requests. It tells you which tools the models mix up, how often, and why.

toollens run -s "python server.py" -t tasks.jsonl -m ollama:qwen3:4b

01 · Inspect

Connects to any MCP server.

Over stdio or Streamable HTTP. It snapshots every tool the server exposes. Here that's a 10-tool calendar server.

02 · Lint

Flags look-alikes before any model runs.

Similar tool pairs, missing "use this when / not for" guidance, undocumented parameters and oversized schemas. Eight rules, TL000 to TL007.

TL003 look-alike pairTL004 no guidanceTL001 terse

03 · Generate

Writes the requests that trip models up.

Direct requests, near-misses that lure toward a sibling tool, and out-of-scope requests where the right answer is to call nothing.

70 direct30 near-miss12 no-tool

04 · Validate & run

Same fixed set, every model.

Local runtimes, cloud APIs and an offline BM25 router all face the same validated tasks, so scores stay comparable over time. Nothing gets executed. It scores the decision.

05 · Score

Shows exactly where it breaks.

Rows are the tool a request needed, columns the tool the model picked. Amber is confusion. Grey is a missed call.

26%tool accuracy
terse descriptions

06 · Rewrite

Same tools, better words.

Only the descriptions changed. Accuracy nearly doubled, and the matrix shows the trade-off: richer text also pulled in more false calls.

49%accuracy
6missed calls
8false calls

07 · Gate

Then it guards every pull request.

One flag turns a run into a CI check, and a report diff shows what a description change actually did.

toollens run … --fail-under 0.9 --html report.html
Scroll to assemble

02What I built

Now it's a number I can track, and a check that fails the build.

I rewrite a description, rerun the same fixed tasks, and see exactly which mix-ups went away and which new ones appeared.

112hand-reviewed tasks in the bundled dataset
26% → 49%tool accuracy after rewriting descriptions only
8static lint rules that run before any model
10+model runtimes behind two wire formats

About these numbers. They come from the offline BM25 router, which rewards shared vocabulary, so they overstate how much wording matters to a capable LLM. The demo shows the workflow. Run your own model to get the number that matters for your stack.

02 · Market data · 2026

IPO Radar

Every open Indian IPO, read at a glance, twice a day.

What
A tracker and a read-only facts API for Indian IPOs
Data
Gathered from multiple sources, with fallbacks and self-healing
Stack
Python, SQL, OpenAPI 3, server-rendered pages
Ships as
Email digest, web app, REST API (beta)

01The problem I hit

One IPO's numbers lived on three pages, and they rarely agreed.

The schedule was in one place, the grey market premium in another, live subscription in a third. Names were spelled differently and dates came without a year. And one source changed its layout four times in a single week.

The idea

One colour per IPO.

Twice a day it collects the facts on every open IPO and gives each one a colour, healthiest first, in a single email.

strongmixedweak

01 · Gather

Several sources, each with a fallback.

Schedule, premium and subscription each come from their own source. A slow source gets retried, and a missing figure reads as unknown, never as zero.

02 · Self-heal

When a source changes shape, it adapts.

Columns are found by their headers. When the headers change, they're inferred from the numbers themselves and only accepted if the arithmetic holds. The drift still gets reported.

listing = price + premium · % = premium ÷ price

03 · Join

Matched on price, then name.

Each IPO is joined across sources and cross-checked three ways. An ambiguous match is refused, so no company inherits another's numbers.

price bandstatuspercent within 1pp

04 · Score

Four signals, one colour.

Premium level, premium trend, institutional demand and issue type, scored against thresholds each reader sets for themselves.

≥ 4strong
2–3mixed
≤ 1weak

05 · Remember

Every reading feeds tomorrow's trend.

Each run is logged. The trend is this reading minus the last one, and the log is how the scoring will eventually be checked against real listings.

06 · Notify

Quiet unless something changed.

One email per reader, covering the whole board. If nothing moved since the last run, nobody hears from it.

07 · Serve

And the same facts, as an API.

Read-only, versioned and keyed. Responses come from the latest run and are never computed on request, so API traffic can't reach the sources.

GET /v1/ipos?status=open
Scroll to assemble

02What I built

Now a layout change is a line in the log, not a wrong number.

Sources still change without warning. IPO Radar reads through the change and flags it, and no reader's inbox ever shows it.

03The facts API

Facts in, facts out.

  • 01Facts only.Raw figures, never the score or colour, so nobody mistakes a heuristic for data.
  • 02null is not 0."Not published yet" and "published, and zero" are different answers.
  • 03Reviewed keys, hard limits.A person approves every key. 200 calls a day and 2 a second, each.
  • 04Stable contract.Versioned, with an OpenAPI spec. Fields get added, never renamed.
GET /v1/ipos?status=open
x-api-key: ••••••••••••

200 OK
{
  "as_of": "2026-09-01T15:00:12+05:30",
  "check_type": "final",
  "count": 1,
  "ipos": [{
    "id": "esds-software-solution",
    "type": "mainboard",
    "status": "open",
    "issue_price": 429.0,
    "gmp_percent": 58.28,
    "qib": 272.19,
    "listing_price": null
  }]
}
Trimmed sample from the docs. Every field is always present, and unknown values are null.
3sources, joined and cross-checked on every run
6past source layouts, all read identically
2×checks a day, with an early layout check before the first
200API calls a day per reviewed key

Not investment advice. IPO Radar reports data against thresholds each reader sets, and never tells anyone what to do. Grey market premiums are unofficial, and the scoring is a heuristic that hasn't been statistically validated yet. That's what the log is for.

NextSay hello

Building something similar?