A working portfolio

Systems I have built, briefly described and visually indexed

A short tour of the live and private builds behind the orchestrator that produced this morning's brief. Each is a system in its own right; each has shipped to a real user; each was built to a regulator-defensible standard. Status pills mark whether the build is publicly accessible, privately deployed, or available as a working demonstration on request.

Private koda command centre vault · 247 memory · layered agents · 8 running crons · 16 sessions · 4 active ● online £42/mo memory architecture canonical anchors named, hand-curated, ~50 facts · O(ms) lookup 100% recall BM25 + semantic hybrid vault search · 850 notes · re-ranked 68% recall BM25 fallback disk scan · last resort · cold-start 45% recall
Build I

Koda — an Anthropic-native command centre

Claude Code as the interface, MCP for tools, LiteLLM as the model gateway, Langfuse for observability, a layered vault with three retrieval backends, and a launchd auto-committer as the sole git writer so concurrent sessions don't race each other. The interesting bit is the operating discipline: every memory cites its source file, the vault is treated as if it were going to be audited, and the system refuses to invent facts it cannot point at.

Stack · Claude · MCP · LiteLLM · Langfuse · self-hosted Built · 2026
Private chairman synthesis strategist analyst counsel historian red team technologist regulator ethicist 8 personas · 1 chairman · adversarial red-team pass
Build II

Council architecture

A reusable Anthropic-native stress-test pattern. Six expert personas read the same evidence in parallel; a chairman synthesises; a red team attacks the synthesis; a final pass closes the loop. Non-negotiable quality bars: every claim cited, named individuals with titles, no fabrication. Built once for a strategic decision; reusable on anything that needs adversarial scrutiny — opposing-counsel reads, deal-room scrutiny, novel-issue advice memos, vendor bake-offs.

Pattern · multi-agent orchestrator Used on · 12+ decisions
Live @dk_ops_bot online · owner allowlisted log payment ₦185k from Adamu, suit deposit logged · ledger updated CRM client_id 0042 ✓ whose suit needs final fit? this week 3 clients · attached list tap to confirm reminders agent loop 1 · intent 2 · tool call 3 · ledger write 4 · confirm 5 · audit log
Build III

Kalu — autonomous ops agent

The operations agent for an SME tailoring client (Doti Kau, Abuja). Telegram-native, owner-allowlisted, multimodal. Takes shorthand commands ("log payment ₦185k from Adamu, suit deposit"), writes to the live ledger, and replies with structured confirmations. Owns its domain end-to-end. Built on Anthropic's Claude with an OpenClaw sandboxed runtime and a strict citation-required output policy.

Stack · Claude · OpenClaw · Telegram Status · production
Demo on request portfolio drift · 30d actual target IPS Sable · recommendation Rebalance gilts → short-duration Drift: +4.2% from IPS Curve: 10Y vs 2Y inverted 14 bp Tax: capital-gains headroom £8,400 Reg: TS-compliant (PS25/22) Sources BoE_yieldcurve_2026-05-26.csv IPS_v3.pdf · §4.2 interaction log · hash-chained 7a3f…91e2 → 4c1b…8a73 → 9d2e…f0a4 → b6c5…2e18 every output reviewable; nothing auto-executed
Build IV

Sable — a regulator-defensible advisory agent

A Claude-Sonnet advisory agent producing portfolio recommendations from real-time market data and a written investment policy statement. Every output is reviewable, every claim is cited to a source file, every interaction is logged in a hash-chained tamper-evident audit trail. Built to test what a fully governable consumer-facing advisor would look like under the FCA's PS25/22 Targeted Support regime. Has its own published training programme and failure modes.

Stack · Claude · IPS-grounded · hash-chained Status · staging
Live Pack 3 · Q14 of 50 · 24:18 remaining CIMA BA2 · cost classification A firm classifies its factory rent as a step-fixed cost. Which statement best describes a step-fixed cost? A · varies with each unit of output ✓ B · fixed within a range, jumps at a threshold C · fixed regardless of output D · varies proportionally with output readiness 73 / 100 weak areas · FX risk · 41% · variance · 58% · activity-cost · 62% pace 28 s/Q (target 36)
Build V

BA2 — a data-driven exam engine

A CIMA BA2 mock-exam engine built for a family member sitting the paper. 250-question bank across five packs, three question types, exam-weighted /150 scoring, three modes (Exam, Learn, Full-Exam-Conditions), and a hub-side analytics layer surfacing readiness, weak areas, confidence drift and pace. Two-day build, end-to-end. Live now at abiru.ai/nafisa in time for the exam window.

Live at · abiru.ai/nafisa Built in · 48 hours
Private net position £ 247,810 ▲ +£4,210 · 30d allocation cash 42% equity 31% property 18% gilts 9% Sable · daily memo Your cash buffer covers 9 months of base expenses. Next FY tax cliff (ISA cap) is 312 days out — no action this week. Two unusual transactions flagged for review (one Zubaydah transfer; one Nigerian remittance). Open for confirmation. 15 tools · 1,700+ transactions · audited
Build VI

WealthOS — a personal financial life dashboard

A private financial dashboard with bank feeds, portfolio analytics, tax planning, and a Sable-backed daily memo that surfaces only what genuinely needs attention. 5,000+ lines, 1,700+ transactions, compliance-grade audit trails. Built first for myself; the design lessons (silent refresh, single daily memo, mobile-primary, every action reviewable) have become load-bearing constraints for any consumer-facing build that follows.

Stack · Claude · bank feeds · 15 tools For · self
Private draft — templated identical boilerplate detected in 3 of 3 letters rewritten — three voices distinct voice · 230 words · ready to sign cross-check vs atty's 39-item list 37 / 39 ready 2 lodge-blockers surfaced
Build VII

Letter-pack rewriter + checklist cross-check

This week's craft piece. Three "independent" support letters came back templated — identical impact paragraph, identical closing block. A caseworker reading them back-to-back would clock the templating immediately. The system rebuilt them in three distinct voices and cross-checked an outstanding-document tracker against the attorney's 39-item formal list — surfacing two lodge-blockers we had mis-marked complete. The cross-check pattern is portable to deal closings, transaction checklists, regulator submissions.

Pattern · independence-detect + audit cross-check Built in · 48 hours
Live Doti Kau OS dashboard
Build VIII

Doti Kau OS — a full tailoring business operating system

Nine modules, sixteen tables, additive pricing, bill-of-labour model. Built for an Abuja bespoke menswear business across the entire commission lifecycle — client intake, fitting, fabric, pattern, cut, sew, finish, delivery, follow-up — with a financial HQ on top, a tailor-pay engine in the middle, and a CRM at the front. The forensic audit pass during build surfaced an inventory plug equivalent to ₦230.5M against retained earnings. Live, in production, currently used by the team. Companion agent (Kalu) sits on top of this system.

Modules · 9 Tables · 16 Companion · Kalu agent Status · production
The above are working systems, not concepts. Each can be walked through live on request.
Confidential · Prepared for Mark Dean · 28 May 2026