Org card: lead with the consultancy, then the inference work

README.md

@@ -9,56 +9,74 @@ pinned: false
 
 # SG Systems
 
-On-device inference for Apple Silicon. We work on the two constraints that
-actually decide whether a model runs on a laptop or a phone:
+**Consulting & forward-deployed engineering for Southeast Asian applied AI.**
 
-**Capacity** — a 7B model at fp16 is 15.5 GB; Apple platforms terminate a
-process near ~3.7 GB. Closing a 4x gap means deciding which weights can be
-destroyed and which cannot.
+We embed senior engineers with client teams and ship working systems in weeks
+rather than producing strategy decks. Fixed scope, outcome-based pricing, no
+hourly billing.
 
-**Bandwidth** — once the weights fit, throughput depends on whether the compute
-maps onto the hardware at all. Sequential state-space recurrences do not map
-onto matrix units by default, and mobile compilers fall back to scalar paths.
+Serving banking, insurance, fintech, healthcare, retail, government and
+mid-market commerce across SEA — where a deployment has to be PDPA-compliant,
+locally tuned, and native to the channels people actually use.
 
-Those are separate problems, and solving one does nothing for the other.
+[sg-systems.co](https://www.sg-systems.co/)
+
+## Why we publish inference work
+
+Regional deployment runs into two constraints that a general-purpose API does
+not solve.
+
+**Data residency.** PDPA and its neighbours make "send the customer's data to a
+US endpoint" a compliance conversation before it is an engineering one.
+Inference that never leaves the device sidesteps that conversation entirely.
+
+**Hardware reality.** SEA deployments run on the devices people have, not on
+datacentre accelerators. A model that needs 15 GB is not a product; one that
+fits in 3.7 GB is.
+
+So we work on making capable models run locally. What we publish here are the
+artifacts and measurements from that work.
 
 ## Projects
 
-**FORGE** — mixed-precision post-training quantization. Not uniform low-bit:
-the result depends less on the average bit width than on which tensor families
-are excluded from it. Crushing a Mamba-2 state projection to ternary produces a
-model that is fluent and confidently wrong.
+**FORGE** — mixed-precision post-training quantization. The result depends less
+on the average bit width than on which tensor families are excluded from it:
+crushing a Mamba-2 state projection to ternary yields a model that is fluent and
+confidently wrong.
 
 **HELIX** — a Metal prefill kernel for the SSM scan in Mamba-2 hybrids,
 integrated into `llama.cpp` through a 42-line patch. Chunk-parallel
-decomposition; ~3x on the isolated op for the shapes upstream's fast path
-declines.
+decomposition; ~3x on the isolated op for shapes upstream's fast path declines.
+
+Source: [github.com/sg-systems-co/forge-helix](https://github.com/sg-systems-co/forge-helix)
 
 ## Models
 
 | | |
 |---|---|
-| [`Falcon-H1-7B-FORGE-v2`](https://huggingface.co/sgsystems/Falcon-H1-7B-FORGE-v2) | `Falcon-H1-7B-Instruct` at 2.06 bpw / 3.48 GB. Runs on Apple Silicon in 3.73 GB resident at 75 tok/s decode. |
+| [`Falcon-H1-7B-FORGE-v2`](https://huggingface.co/sgsystems/Falcon-H1-7B-FORGE-v2) | `Falcon-H1-7B-Instruct` at 2.06 bpw / 3.48 GB. Runs on Apple Silicon in 3.73 GB resident at 75 tok/s decode. **English only** — this is an engineering artifact from the compression work, not a SEA-language model. |
 
-**Read the model card before using it.** This one needs `repeat_last_n = 2048`
-— llama.cpp's default of 64 is too small to suppress the cross-turn repetition
-these models fall into, and shipping the usual sampling triple without it
-reproduces the bug.
+**Read the model card before using it.** It needs `repeat_last_n = 2048`;
+llama.cpp's default of 64 is too small to suppress the cross-turn repetition
+these models fall into, and the usual sampling triple without it reproduces the
+bug.
 
 ## How we report numbers
 
-Measurements here come with the conditions attached, and with the cases that
-did not work.
-
-Some things we published that were not what we first assumed:
+Measurements come with their conditions attached, and with the cases that did
+not work. Things we published that contradicted our own assumptions:
 
 - On M5, `simdgroup_matrix` at bf16 is **slower** than at fp32 (10.88 vs 16.12
   TFLOP/s). It is not the matrix hardware. Reduced precision only pays through
   `MetalPerformancePrimitives`, and only with both operands bf16.
 - Adding a repetition penalty did **not** fix our repetition loops. The penalty
-  *window* did, and only above 2048 tokens. Within-turn repetition was 0% the
-  whole time; the failure was cross-turn.
-- The quantized models still lose meta-instruction following. Direct
-  instructions work; "repeat my first message" does not.
+  *window* did, and only above 2048 tokens. Within-turn repetition was 0%
+  throughout; the failure was entirely cross-turn.
+- The quantized model still loses meta-instruction following, and no perplexity
+  comparison against fp16 has been run.
 
 Where a claim is not measured, we say so rather than rounding it up.
+
+## Contact
+
+sebastian@sg-systems.co