Abdullah Siddiki Applied AI
Open to work · remote or NYC area (hybrid / in person)
01MaktabSchool administration platform for Islamic schools. Admissions, Stripe tuition payments, academics and FERPA-audited student records.375API operations3,460automated tests23Stripe event types
02Open Deep ResearchMulti-agent research API evaluated on 20 DeepResearch Bench tasks. The eval traced report quality to one reranking stage.0.847FACT accuracyr 0.69recall vs RACE$0.22median cost
03jev-cc-analyzerAnswers 16 analyst questions about an earnings call with a decision model. Evaluated for accuracy, calibration and injection robustness.157 ms16 decisions91%acceptable vs gold0.10calibration error
05candidsQR-code photo galleries for weddings and parties on an all-Cloudflare stack. Image moderation runs on a decision model picked by benchmark.0.866AUROC, Jev 1.1410,146images scored865tests
06KhaldunMulti-agent equity research on Agno. Eight analyst agents feed a synthesizer, and a permission-aware chat answers over reports and team conversations.8analyst agents1typed LLM call each3chat sourcesA multi-agent equity research system. Eight analyst agents compute their metrics in Python and make one typed LLM call each. A synthesizer writes the brief, and a chat answers questions over reports and conversations, including teammates’ shared conversations, within each user’s permissions.
SynthesizerPrototype
Receives eight typed summaries and writes the brief: executive summary, key insights, unified risks, thesis, catalysts, and a BUY, HOLD or SELL call with a confidence level. Hover an analyst to see what it reads, computes and writes.
built designed
01Agents
Every analyst follows the same five-file contract. Data intake and math are plain Python, so every number is deterministic and unit-testable. The LLM receives a typed payload of computed facts and returns a typed narrative.
models.pyPydantic contracts for input, LLM input, LLM output and the dashboard JSON.
data_intake.pyAsync Alpha Vantage prefetch with asyncio.gather. No tools.
analysis.pyPure Python math. No LLM calls.
agent.pyAn Agno Agent with input_schema, output_schema and tools=[].
routes.pyA FastAPI router to run and stream one analyst alone.
class FundamentalsAgent(Agent): def __init__(self): super().__init__( name="Fundamentals Sage", model=OpenRouter(id="google/gemini-2.5-flash"), tools=[], input_schema=FundamentalsSynthesisInput, output_schema=FundamentalsSynthesisOutput, use_json_mode=True, )
The first build had Fundamentals Sage computing P/E, PEG, ROE, ROIC, net debt to EBITDA, free cash flow, CAPM, WACC and a five-year DCF before the model saw anything. The model wrote takeaways, risks and chart notes from those numbers.
Keeping tools out of the analysis path makes each analyst one predictable LLM call. Tools are reserved for the chat agents, where questions are open-ended.
| Analyst | Data | Python computes | Status |
|---|---|---|---|
| Fundamentals Sage | OVERVIEW, income statement, balance sheet, cash flow, daily prices, 10-year Treasury | P/E, PEG, ROE, ROIC, net debt/EBITDA, free cash flow, CAPM, WACC, 5-year DCF | Built |
| Sentiment Pulse | NEWS_SENTIMENT, fetched in weekly chunks and deduplicated by URL | Daily score series, 7- vs 14-day z-score, velocity, coverage intensity, surprise index | Built |
| Analyst Consensus | EARNINGS_ESTIMATES, daily prices | Revision delta and velocity, estimate dispersion, estimate-versus-price divergence | Designed |
| Earnings Oracle | EARNINGS, EARNINGS_ESTIMATES, earnings call transcripts | EPS and revenue surprise, extracted guidance, hedge-word and polarity counts | Designed |
| Market Context Navigator | SECTOR, ETF_PROFILE, index prices, Treasury yields, CPI, unemployment, GDP | Return correlations, relative strength versus sector, risk-on or risk-off regime | Designed |
| Technical Patterns Scout | Daily prices, RSI, MACD, ADX, Bollinger bands, OBV, options chains | Fractal pivots, volume by price, breakouts on 1.5× volume, unusual options flow | Designed |
| Risk Sentinel | INSIDER_TRANSACTIONS, LISTING_STATUS, daily prices | Clustered insider selling, 1-day 95% VaR, worst-5-day stress replay, risk tags | Designed |
| News Velocity Tracker | NEWS_SENTIMENT, TOP_GAINERS_LOSERS | Articles per day z-scores, sentiment volume, impact scores, catalyst tags | Designed |
02Architecture
The analysis path is a deterministic workflow. The chat path is a routing team over a permission-aware index. Identity, roles and sharing come from WorkOS.
Hover or tap any component to trace its connections, or pick a flow to step through a request end to end.
03Orchestration
The two paths have different shapes, so they use different Agno primitives.
Analysis path
A fixed fan-out and fan-in: one Parallel block of eight steps, then the synthesizer, then the indexer. No coordinator LLM decides who runs, every analyst keeps its typed schema, and each step gets retries and a timeout.
skip_on_failure turns a failed analyst into a degraded lane the synthesizer must name.get_step_content("analysts"). The first build read previous_step_content, which after a Parallel block is a summary string.Chat path
Turns are open-ended, so a team routes each one to a single specialist: the Report QA agent for questions about past analysis, or the live-data agent for current numbers. Route mode avoids a coordinator loop that would add latency and cost to every turn.
04Retrieval
A user can ask about their own reports and conversations and about conversations a teammate has shared. Access is decided by the server from the verified token. The model only sees what the user is allowed to read.
| Source | Retrieval unit | Tagged with | Indexed |
|---|---|---|---|
| Briefs | One chunk per brief section | org, owner, run, symbol, as-of, visibility | End of run |
| Analyst outputs | Text sections and single-metric facts | Above, plus agent and JSON pointer | End of run |
| Own conversations | Question and answer pairs | conversation, owner, visibility | After each turn, async |
| Shared conversations | Same pairs | Above, plus author | Visible once shared |
The JWT carries user, org and role. AgentOS JWTMiddleware maps roles to scopes such as runs:create and sessions:read.
Owner, org-member and explicit-share relations. The retriever lists readable objects once per turn and caches them for the session.
Hybrid search and the access predicate run in one Postgres query, so nothing outside the readable set is ever scored.
Agentic filters can narrow by symbol, date or source type. Org, owner and visibility come only from the token.
Answers cite JSON pointers and message IDs. A validator rejects any citation that is not in the retrieved set.
News articles and teammates’ messages are wrapped as data, never as instructions, before they reach another user’s chat.
05Interface
Khaldun had no front end, so I designed one for it. The surfaces are dark malachite with ivory ink and brass for actions. Each analyst gets one hue, and every claim in the UI sits next to its source.
Surfaces and ink
#081713#0c1f1a#11291f#163126#1f3d32#eeeadf#d4a64aOne hue per analyst
#7fc8a9#f28e6b#8fb4ff#e3c567#6fd0e0#c49bf5#f06a7a#b7d36bCalls and type
#5fc79a#b9c2ba#e66a6aIBM Plex Sans 600
Body copy in Plex Sans at 15 px. Sentence case everywhere.
PLEX MONO · 31.4× · AS OF 2026-10-22
Strong free cash flow and a clean balance sheet, offset by slowing estimate revisions and a multiple above its five-year range.
fundamentals · dcf.margin_of_safetyconsensus · revisions.velocity_30dRisk Sentinel was rate limited. Risks below come from 7 of 8 analysts.
Mock data on a fictional ticker. The interface has not been built.
06Evals
The evaluation approach comes from my other projects: grade retrieval and grounding separately, validate any judge before trusting it, and treat a permission leak as a failing build.
| Eval | Measures | Gate |
|---|---|---|
| Retrieval golden set | Recall@10 and MRR per source type on questions with labelled chunks | Regression against the last release |
| Citation grounding | Share of cited claims supported by the cited chunk, judged per claim | Regression against the last release |
| Cross-tenant leakage | Chunks from another org or an unshared conversation in any result set | Must be zero |
| Routing | Turns sent to the correct specialist | Tracked per release |
| Freshness | Answers citing expired data without a warning | Must be zero |
07Decisions
ADR 01
The analysis is a fixed fan-out and fan-in. A workflow is deterministic, has no coordinator LLM cost, and keeps every analyst’s schema intact.
Trade-off Adding an analyst is a code change, not a prompt change.
ADR 02
Numbers stay deterministic and testable, and each analyst costs one predictable LLM call.
Trade-off Analysts cannot fetch anything they were not given up front.
ADR 03
The access predicate, hybrid search and session data need to meet in one query and one transaction.
Trade-off A Postgres instance to operate.
ADR 04
Access must come from the verified token. In Agno 2.0.5 metadata filters are AND-only, so “mine or shared with me” needs custom SQL anyway.
Trade-off More code than the built-in search tool.
ADR 05
Org, team and per-conversation shares are relationship checks, and the same model can later cover reports and workspaces.
Trade-off A network call per turn, cached per session.
ADR 06
A pair is the unit people ask about, and it keeps attribution and visibility exact.
Trade-off Context that spans many turns is split across chunks.
08Failure modes
| What breaks | Guard | How I'd know |
|---|---|---|
| An analyst fails or times out | skip_on_failure, a degraded flag, and a synthesizer that must name missing lanes. | Degraded badge in the run view; degraded outputs are not indexed. |
| Upstream returns no data | Validate before the LLM call. An empty news feed returns “insufficient data” without a model call. | Per-lane record counts on each trace. |
| Alpha Vantage rate limits | A shared client with per-endpoint TTLs and minute and daily counters. | Cache hit rate and quota counters per run. |
| One user sees another org’s data | Access predicate in SQL from the verified token; no unscoped history tools. | Leakage eval at zero in CI. |
| Injected instructions in news or a shared message | Fenced untrusted context and citation validation. | Red-team cases in the eval suite. |
| Stale numbers presented as current | As-of on every fact and a re-run offer. | Freshness eval. |
09Status
Sep 2025
Fundamentals Sage and Sentiment Pulse, a parallel workflow, a synthesizer prototype, and Postgres with pgvector for sessions.
Dec 2025 to Mar 2026
A shared Alpha Vantage client with caching, retries and quota counters. Sentiment metrics moved fully into Python.
This design
The full workflow, the synthesizer, and chat over reports and team conversations on Agno 3, using AgentOS RBAC with WorkOS, user-scoped knowledge and tracing.
10Stack
Agents
Models
Identity
Data
API
Ops