Systems, and what they were judged on

A three tier classification cascade for job taxonomy

ProblemFree text job titles in five languages had to resolve to a controlled taxonomy. A single LLM call per title was accurate enough and far too expensive to run at catalogue scale.

ApproachThree tiers, cheapest first. A cache fronted by a Bloom filter answers anything seen before. A keyword engine covering eight domains handles the long tail of predictable titles. Only what survives both reaches the model. A review queue catches low confidence output before it enters the taxonomy, and a separate agent path runs a three phase plan of database lookup, then trend and URL enrichment in parallel, then search, then synthesis.

• 03 Ground
• 04 Assemble
• 05 Orchestrate
• 06 Prove
• 07 Serve
• 08 Govern

Tier 1 cache, P50

48 ms

Tier 2 keyword, P50

180 ms

Tier 2 accuracy

81.6%

Tier 3 model, P50

1.2 s

Tier 3 accuracy

95.0%

Languages

5

Bulk import ceiling

10k rows

• Next.js
• TypeScript
• PostgreSQL
• Gemini 2.5 Flash
• Zod
• Vitest

Contract risk and renewal agents inside the office suite

ProblemProcurement contracts sat in a document library nobody read end to end. Penalty clauses, unilateral termination rights and price escalation terms were discovered after they bit.

ApproachTwo declarative agents grounded on the existing document library and list, so nothing left the tenancy and no new system had to be adopted. The first reads a contract and scores each clause across delay and performance penalties, indemnity, termination notice, payment maturity and currency exposure. The second tracks end dates for services, licences and rentals and issues reminders on a fixed schedule.

• 01 Frame
• 02 Source
• 03 Ground
• 08 Govern

Agents

2

Risk tiers

3

Clause categories scored

5

New systems introduced

0

• Microsoft 365 Copilot
• Declarative agents
• SharePoint
• Teams Toolkit

Speech recognition on telephone grade audio

ProblemCall audio arrived narrowband and heavily compressed. Off the shelf transcription was tuned for clean wideband speech and degraded badly on it.

ApproachA pipeline that treats the codec as the primary variable. Raw captures are held alongside every transcript so a regression can be replayed against the original bytes rather than against a re encoded copy, and a smoke set runs on each change.

• 02 Source
• 06 Prove
• 07 Serve

Audio path

GSM narrowband

Artefacts kept per call

raw, wav, transcript

• Azure Speech
• Python
• Shell

Schema compaction for text to SQL

ProblemThe full database schema did not fit usefully in the prompt. Sending all of it crowded out the question; sending a guess at the relevant part broke joins.

ApproachA compaction step that rewrites the schema into a compact representation carrying only the tables, keys and relationships a query can reach, with a test suite of known queries to catch the compaction dropping something a join needed.

• 04 Assemble
• 06 Prove

Prompt component

schema, compacted

Regression suite

SQL tests

• Python
• YAML
• SQL

Token accounting for agent systems

ProblemAgent setups are quoted in capabilities and paid for in tokens. The fixed prefix cost of a harness, per turn, was invisible to the people choosing between harnesses.

ApproachA measurement instrument rather than a product. Formalise the occupancy of a context window, name the symbols, and compute what a given configuration costs before it does any work.

• 01 Frame
• 04 Assemble
• 07 Serve

Symbols catalogued

150

Source files indexed

56

• TypeScript
• LaTeX
• Python