Portfolio dashboard
Positions, exposure and attribution in one view, with drill-through from a headline number to the trades behind it.
Product engineering
Kilwa
Market data, sentiment and research in one investment workspace.
An AI investment-intelligence platform that pulls market data, news sentiment and analyst research into a single workspace, so an investment team can go from a question to a defensible answer without leaving the screen.
- Sector
- Investment & Asset Management
- Year
- 2025
- Duration
- 9 months
- Team
- 7 people
- Platforms
- Web · iOS companion
- Status
- In production
$2.41M
Portfolio under analysis
Live in the reference deployment
40+
Data sources unified
Market, macro, filings and news
100%
Cited outputs
Every generated claim links to a source
3.2s
Median answer time
Retrieval through to rendered response
01 — The problem
A mid-sized asset manager was running its process across four terminals, two data vendors, a shared drive of PDFs and a group chat. Analysts spent more of the week assembling context than forming a view.
They had tried generic AI assistants and abandoned them. The models produced fluent summaries with no lineage, and an analyst cannot put a number in front of an investment committee when they cannot say where it came from.
02 — What we did
We treated citation as a hard product constraint rather than a feature. Nothing surfaces in Kilwa without a source object attached: a price series, a filing paragraph, a dated news item. If the system cannot cite it, the system does not say it.
The research assistant was built as a retrieval-first pipeline. Questions are decomposed into structured sub-queries against the market store and the document index, results are ranked, and only then does a model write prose over material the analyst can click straight through to.
An answer an analyst cannot trace is not an answer. It is a liability.
Interface
What it looks like in use
Two of the screens that carry the most weight in daily use, rebuilt here from the production design system.
Portfolio · Growth MandateLIVE
$2.41MUNDER ANALYSIS
+3.2% today
Sentiment · Semiconductors+0.42
Macro · 10Y real yield1.88%
NAV40 sources
Portfolio, citedEvery headline number drills through to the trades and series points that produced it. Nothing on this screen is unsourced.
Why did margins compress in Q3?4 SOURCES
Input costs rose 210bps year on year10-Q p.14
Pricing held flat through the quarterCALL 11:04
Mix shifted toward lower-margin OEM10-Q p.22
Unsupported claim · rejected by validatorBLOCKED
Retrieval 1.1sGeneration 2.1s
Citation is a hard constraintThe assistant answers only from retrieved material. A sentence that cannot be sourced is rejected before it renders.
Capabilities
What the system does
Grouped by the job each set of capabilities exists to do, rather than by which team built it.
Investment workspace
01 · The daily surfaceReal-time indicators
Streaming prices and macro series with configurable alerting per instrument and threshold.
Watchlists and theses
An analyst pins a thesis to a name and the workspace tracks evidence for and against it over time.
Scenario comparison
Side-by-side modelling of assumptions with the deltas made explicit rather than buried in a spreadsheet.
Research assistant
02 · Retrieval-first, never freehandGrounded question answering
Questions decompose into structured retrieval before any generation happens, so answers are assembled from sources rather than recalled.
Citation-linked responses
Every sentence carries its provenance. Clicking a claim opens the exact passage or series point it rests on.
Filing and document search
Semantic search across annual reports, transcripts and internal notes, scoped by entity and date.
Comparative briefs
Automatic side-by-side briefs across a peer set, built from the same cited material.
Signal and sentiment
03 · Context, quantifiedSentiment analysis
News and transcript sentiment scored per entity with the underlying articles always one click away.
Country and market analysis
Macro dashboards per market with the indicator history and revision trail intact.
Anomaly surfacing
Movement that breaks an instrument's own historical pattern is raised rather than waiting to be noticed.
Insight digest
A morning brief assembled from overnight movement across everything the desk holds or watches.
Architecture
How it is put together
Layer by layer, with the reason each one exists — because the reason is usually the interesting part.
01Ingestion
Scheduled and streaming connectors normalise vendor feeds, filings and news into a common entity model with revision history.
- Airflow
- Kafka
- Entity resolution
02Stores
A time-series store for market data alongside a vector and keyword index for documents, queried together rather than separately.
- TimescaleDB
- pgvector
- OpenSearch
03Retrieval layer
Hybrid dense and lexical retrieval with reranking, returning source objects that the UI can render and link.
- Hybrid search
- Cross-encoder rerank
04Generation
Constrained generation over retrieved passages only, with a citation validator that rejects any unsupported sentence.
- Claude
- Citation validator
05Application
A streaming React workspace where answers render progressively with their citations attached.
- React
- TypeScript
- SSE
Technology
Frontend
- React
- TypeScript
- Vite
- Visx
- TanStack Query
Backend
- Python
- FastAPI
- Celery
- PostgreSQL
Data & AI
- TimescaleDB
- pgvector
- OpenSearch
- Claude API
Platform
- Kubernetes
- Terraform
- Grafana
- Sentry
Outcome
What changed
Measured against how the operation ran before, not against a benchmark chosen after the fact.
-
Every generated claim is traceable. Analysts can take output into an investment committee and defend it line by line.
-
Four tools collapsed into one workspace. Context assembly stopped being the bulk of the working week.
-
Sentiment became evidence, not vibes. Scores always open onto the articles and passages that produced them.
-
Latency held under load. Median answer time stayed near three seconds as the document corpus grew past a million passages.
How it ran
-
Weeks 1-3
Process mapping
Traced how three analysts actually built a view, from first question to committee memo.
-
Weeks 4-8
Data foundation
Entity model, ingestion connectors and the combined time-series and document stores.
-
Weeks 9-18
Retrieval and citation
Hybrid retrieval, reranking and the citation validator that gates generation.
-
Weeks 19-30
Workspace
Dashboard, watchlists, scenario comparison and the streaming assistant surface.
-
Weeks 31-38
Hardening
Load testing against a full corpus, access controls and audit logging for regulated review.
Related work
From the same practice
Recycling & E-Waste
Razor ERP
One connected system for a warehouse that ran on paper.
A mobile-first ERP for warehouse and e-waste operations that replaced a decade of spreadsheets, paper travellers and radio calls with a single connected system that works on the floor..
2024 · 7 months
HR Tech & Recruitment
The Walt
Where candidates move, and where they stall.
A recruitment analytics platform that turns an applicant tracking system's raw event log into a picture of where a hiring funnel actually leaks, with the bottleneck named rather than left to be inferred..
2024 · 5 months
Healthcare & Clinical Communication
Lyfe Languages
Clinical translation where a wrong word has consequences.
A translation platform built specifically for clinical settings, where general-purpose translation is not safe enough and a verified medical dictionary sits between the model and the patient..
2023 · 8 months
Something similar?
Running into the same problem Kilwa solved?
Tell us what you run. We will reply within two business days with what we would build for your situation — and, just as usefully, what we would leave out.