What this engagement is evidence of
Real time data engineering under a latency budget in a regulated market, where the numbers have to reconcile against the source before anyone will act on them. Mostly a platform problem, which is what most production AI turns out to be.
Asset management is a business where timing is everything. SAMCo's analysts were skilled and experienced - but they were making portfolio decisions with information that was, at best, 24 hours old. Overnight batch jobs, Excel risk models maintained by a single analyst, and data silos across trading, compliance, and portfolio systems meant that by the time a risk report reached a portfolio manager's desk, the market had moved. In volatile conditions, a 24-hour lag isn't just inefficiency - it's exposure.
The temptation in projects like this is to build a better dashboard. We spent the first sprint mapping exactly where the latency lived - and most of it was in the data pipeline, not the reporting layer. Overnight batch jobs were running because the previous architecture couldn't support live queries across four data sources simultaneously. We re-architected the data layer with a streaming pipeline before touching a single user-facing component. The natural language interface came last - built on top of a foundation that could actually support real-time responses.
The platform combines a real-time streaming pipeline with a model powered query interface. WebSocket connections maintain live feeds from Bloomberg, the internal trading platform, and the compliance database. A custom risk model - co-designed with SAMCo's head of risk - runs continuously over the live data, updating exposure calculations every 60 seconds. Portfolio managers interact with the system through a natural language interface: they ask questions in plain English and get structured answers with the underlying data exposed for inspection. The system also runs a configurable alert engine that triggers on user-defined risk thresholds.
Streaming data pipeline
Real-time WebSocket connections to Bloomberg API, the internal trading platform, and the compliance database. Position and pricing data updates every 60 seconds. Historical data is preserved in PostgreSQL for trend analysis and backtesting.
AI risk model
A custom risk model trained on SAMCo's portfolio history, co-designed with the head of risk. Runs on live position data to produce VaR estimates, currency exposure summaries, and sector concentration metrics - updating continuously rather than nightly.
Natural language query interface
Model powered query interface grounded in live portfolio data. Queries like 'What is my USD exposure if the dollar moves 2% against INR?' return a structured answer with the calculation methodology, data sources, and confidence range.
Configurable alert engine
Portfolio managers define risk thresholds (e.g. 'alert me if tech sector exposure exceeds 30%') and receive push notifications the moment live data crosses the threshold - replacing the previous end-of-day email digest.
Compliance audit trail
Every AI-generated insight, alert trigger, and query response is logged with the underlying data snapshot, model version, and timestamp. SAMCo's compliance team can reconstruct the exact information state at any moment a decision was made.
The technical metrics tell one story. The operational shift - from reactive risk management to proactive, real-time intelligence - tells a bigger one.
24-hour batch lag eliminated. Portfolio managers save 3 hours per day. Platform handles $2B+ AUM with full audit trails on every AI-generated insight.
Live market data is messy. Bloomberg feeds drop out. Internal systems have maintenance windows. Position data can arrive out of sequence. We spent significant engineering time on fallback states, data staleness indicators, and graceful degradation - so the platform always communicates clearly when data is fresh versus cached, rather than silently serving stale information as if it were live.
Free-form natural language responses in a financial context are a liability. Portfolio managers need structured, scannable answers - not prose. We designed a strict output schema with a direct answer, supporting data table, methodology note, and confidence indicator. This also made it easier to validate outputs programmatically before serving them.
The biggest adoption barrier wasn't portfolio managers - it was the existing analytics team who feared the system made their role redundant. We repositioned their role explicitly as 'model owners and validators': they define the risk parameters, review model outputs weekly, and own the alert threshold configurations. This turned the team from sceptics into champions.
“We used to brief portfolio managers with yesterday's numbers. Now they get real-time answers to questions they didn't even know to ask. The natural language interface has changed how our entire team thinks about data.”
Chief Technology Officer, SAMCo
Model versions are deliberately not listed. The architecture treats the provider as a configuration decision, and a version printed on a page is out of date within the quarter.
engineering / other engagements
Send the task, the data it runs on and the systems it has to touch. You get a written read from an engineer within a working day.